NONLINEARCONTROL
EditedbyMeralAltnay
Contents
Preface IX
Chapter 1 Application of InputOutput Linearization 1
Erdal ehirli and Meral Altinay
Chapter 2 LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter 21
Seigo Sasaki
Chapter 3 Nonlinear Control Applied to the Rheology of
Drops in Elongational Flows with Vorticity 37
Israel Y. Rosas, Marco A. H. Reyes, A. A. Minzoni and E. Geffroy
Chapter 4 Robust Control Research of Chaos Phenomenon
for DieselGenerator Set on Parallel Connection 57
Manlei Huang
Chapter 5 A Robust State Feedback Adaptive
Controller with Improved Transient Tracking Error
Bounds for Plants with Unknown Varying Control Gain 79
A. Rincon, F. Angulo and G. Osorio
Chapter 6 A Robust Motion Tracking
Control of PiezoPositioning
Mechanism with Hysteresis Estimation 99
Amir Farrokh Payam,
Mohammad Javad Yazdanpanah and Morteza Fathipour
Chapter 7 Nonlinear ObserverBased Control Allocation 115
Fang Liao, Jian Liang Wang and KaiYew Lum
Chapter 8 Predictive Function Control of the
SingleLink Manipulator with Flexible Joint 129
Zhihuan Zhang and Chao Hu
VI Contents
Chapter 9 On Optimization Techniques for a
Class of Hybrid Mechanical Systems 147
Vadim Azhmyakov and Arturo Enrique Gil Garca
Chapter 10 Optimized Method for Real Time Nonlinear Control 163
Younes Rafic, Omran Rabih and Rachid Outbib
Chapter 11 Nonlinear Phenomena and Stability
Analysis for Discrete Control Systems 187
Yoshifumi Okuyama
Preface
All practical systems contain nonlinear dynamics. Control system development for
these systems has traditionally been based on linearized system dynamics in
conjunction with linear control techniques. Sometimes it is possible to describe the
operationofsystemsbyalinearmodelarounditsoperatingpoints.Linearizedsystem
can provide approximate behavior of the system. But in analyzing the overall system
behavior, the resulting system model is inadequate or inaccurate. Moreover, the
stability of the system cannot be guaranteed. However, nonlinear control techniques
takeadvantageofthegivennonlineardynamicstoproducehighperformancedesigns.
NonlinearControlSystemsrepresentanewtrendofinvestigationduringthelastfew
decades.Therehasbeengreatexcitementoverthedevelopmentofnewmathematical
techniquesforthecontrolofnonlinearsystems.Methodsfortheanalysisanddesignof
nonlinearcontrolsystemshaveimprovedrapidly.Anumberofnewapproaches,ideas
andresultshaveemergedduringthistime.Thesedevelopmentshavebeenmotivated
bycomprehensiveapplicationssuchasmechatronic,robotics,automotiveandaircraft
controlsystems.
The book is organized into eleven chapters that include nonlinear design topics such
as Feedback Linearization, Lyapunov Based Control, Adaptive Control, Optimal
Control and Robust Control. All chapters discuss different applications that are
basically independent of each other. The book will provide the reader with
information on modern control techniques and results which cover a very wide
application area. Each chapter attempts to demonstrate how one would apply these
techniquestorealworldsystemsthroughbothsimulationsandexperimentalsettings.
Lastly,Iwouldliketothankalltheauthorsfortheirexcellentcontributionsindifferent
applicationsofNonlinearControlTechniques.Despitetherapidadvancesinthefield,
I believe that the examples provided here allow us to look through some main
research tendencies in the upcoming years. I hope the book will be a worthy
contribution to the field of Nonlinear Control, and hopefully it will provide the
readerswithdifferentpointsofviewonthisinterestingbranchofControlEngineering.
Dr.MeralAltnay
KocaeliUniversity,
Turkey
1
Application of InputOutput Linearization
Erdal ehirli and Meral Altinay
Kastamonu University & Kocaeli University
Turkey
1. Introduction
In nature, most of the systems are nonlinear. But, most of them are thought as linear and the
control structures are realized with linear approach. Because, linear control methods are so
strong to define the stability of the systems. However, linear control gives poor results in
large operation range and the effects of hard nonlinearities cannot be derived from linear
methods. Furthermore, designing linear controller, there must not be uncertainties on the
parameters of system model because this causes performance degradation or instability. For
that reasons, the nonlinear control are chosen. Nonlinear control methods also provide
simplicity of the controller (Slotine & Li, 1991).
There are lots of machine in industry. One of the basic one is dc machine. There are two
kinds of dc machines which are brushless and brushed. Brushed type of dc machine needs
more maintenance than the other type due to its brush and commutator. However, the
control of brushless dc motor is more complicated. Whereas, the control of brushed dc
machine is easier than all the other kind of machines. Furthermore, dc machines need to dc
current. This dc current can be supplied by dc source or rectified ac source. Three phase ac
source can provide higher voltage than one phase ac source. When the rectified dc current is
used, the dc machine can generate harmonic distortion and reactive power on grid side.
Also for the speed control, the dc source must be variable. In this paper, dc machine is fed
by three phase voltage source pulse width modulation (PWM) rectifier. This kind of
rectifiers compared to phase controlled rectifiers have lots of advantages such as lower line
currents harmonics, sinusoidal line currents, controllable power factor and dc link voltage.
To make use of these advantages, the filters that are used for grid connection and the control
algorithm must be chosen carefully.
In the literature there are lots of control methods for both voltage source rectifier and dc
machine. References (Ooi et al., 1987; Dixon&Ooi, 1988; Dixon, 1990; Wu et al., 1988, 1991)
realize current control of L filtered PWM rectifier at three phase system. Reference (Blasko
& Kaura, 1997) derives mathematical model of Voltage Source Converter (VSC) in dq and
 frames and also controlled it in dq frames, as in (Bose, 2002; Kazmierkowski et al.,
2002). Reference (Dai et al., 2001) realizes control of L filtered VSC with different decoupling
structures. The design and control of LCL filtered VSC are carried out in dq frames, as in
(Lindgren, 1998; Liserre et al., 2005; Dannehl et al., 2007). Reference (Lee et al., 2000; Lee,
2003) realize inputoutput nonlinear control of L filtered VSC, and also in reference
(Kmrcgil & Kkrer, 1998) Lyapunov based controller is designed for VSC. The feedback
linearization technique for LCL filtered VSC is also presented, as in (Kim & Lee, 2007; Sehirli
Applications of Nonlinear Control 2
& Altnay, 2010). Reference (Holtz, 1994) compares the performance of pulse width
modulation (PWM) techniques which are used for VSC. In (Krishnan, 2001) the control
algorithms, theories and the structure of machines are described. The fuzzy and neural
network controls are applied to dc machine, as in (Bates et al., 1993; Sousa & Bose, 1994).
In this chapter, simulation of dc machine speed control which is fed by three phase voltage
source rectifier under input output linearization nonlinear control, is realized. The speed
control loop is combined with inputoutput linearization nonlinear control. By means of the
simulation, power factor, line currents harmonic distortions and dc machine speed are
presented.
2. Main configuration of VSC
In many industrial applications, it is desired that the rectifiers have the following features;
highunity power factor, low input current harmonic distortion, variable dc output voltage
and occasionally, reversibility. Rectifiers with diodes and thyristors cannot meet most of
these requirements. However, PWM rectifiers can provide these specifications in
comparison with phasecontrolled rectifiers that include diodes and thyristors.
The power circuit of VSC topology shown in Fig.1 is composed of six controlled switches
and input L filters. Acside inputs are ideal threephase symmetrical voltage source, which
are filtered by inductor L and parasitic resistance R, then connected to threephase rectifier
consist of six insulated gate bipolar transistors (IGBTs) and diodes in reversed parallel. The
output is composed of capacitance and resistance.
Fig. 1. L filtered VSC
3. Mathematical model of the VSC
3.1 Model of the VSC in the threephase reference frame
Considering state variables on the circuit of Fig.1 and applying Kirchhoff laws, model of
VSC in the threephase reference frame can be obtained, as in (Wu et al., 1988, 1991; Blasko
& Kaura, 1997).
The model of VSC is carried out under the following assumptions.
The power switches are ideal devices.
All circuit elements are LTI (Linear Time Invariant)
The input AC voltage is a balanced threephase supply.
Application of InputOutput Linearization 3
For the threephase voltage source rectifier, the phase duty cycles are defined as the duty
cycle of the top switch in that phase, i.e., d
a
= d(S
1
), d
b
= d(S
3
), d
c
= d(S
5
) with d representing
duty cycle.
Ji
u
Jt
= 
R
I
i
u
I
dc
_J
u

1
S
J
k
c
k=u
_ +I
u
(1)
Ji
b
Jt
= 
R
I
i
b
 I
dc
_J
b

1
S
J
k
c
k=u
_ +I
b
(2)
Ji
c
Jt
= 
R
I
i
c
I
dc
_J
c

1
S
J
k
c
k=u
_ +I
c
(3)
JI
dc
Jt
=
1
C
(i
u
J
u
+i
b
J
b
+i
c
J
c
) 
1
C
i
dc
(4)
This model in equations (1) (4) is nonlinear and time variant. Using Park Transformation,
the acside quantities can be transformed into rotating dq frame. Therefore, it is possible to
obtain a timeinvariant system model with a lower order.
3.2 Coordinate transformation
On the control of VSC, to make a transformation, there are three coordinates whose relations
are shown by Fig 2, that are abc,  and dq. abc is three phase coordinate,  is
stationary coordinate and dq is rotating coordinate which rotates speed.
Fig. 2. Coordinates diagram of abc,  and dq
Applications of Nonlinear Control 4
The dq transformation is a transformation of coordinates from the threephase stationary
coordinate system to the dq rotating coordinate system. A representation of a vector in any
ndimensional space is accomplished through the product of a transpose ndimensional
vector (base) of coordinate units and a vector representation of the vector, whose elements
are corresponding projections on each coordinate axis, normalized by their unit values. In
three phase (three dimensional) space, it looks like as in (5).
X
ubc
= o
u
b
u
c
u
] _
x
u
x
b
x
c
_ (5)
Assuming a balanced threephase system, a threephase vector representation transforms to
dq vector representation (zeroaxis component is 0) through the transformation matrix T,
defined as in (6).
I =
2
3
_
cos(t) cos(t 
2
3
n) cos(t +
2
3
n)
sin(t) sin(t 
2
3
n) sin(t +
2
3
n)
_ (6)
In (6), is the fundamental frequency of threephase variables. The transformation from X
ubc
(threephase coordinates) to X
dq
(dq rotating coordinates), called Park Transformation, is
obtained through the multiplication of the vector X
ubc
by the matrix T, as in (7).
X
dq
= I. X
ubc
(7)
The inverse transformation matrix (from dq to abc) is defined in (8).
I =
2
3
l
l
l
l
l
cos(t)
cos(t 
2
3
n)
cos(t +
2
3
n)
sin(t +
2
3
n)
sin(t 
2
3
n)
sin(t +
2
3
n)
1
1
1
1
1
(8)
The inverse transformation is calculated as in (9).
X
ubc
= I. X
dq
(9)
3.3 Model of the VSC in the rotating frame
Let x and u be the phase state variable vector and phase input vector in one phase of a
balanced threephase system with the state equation in one phase as in (10).
X
= Ax +Bu (10)
Where A and B are identical for the three phases. Applying dq transformation to the three
phase system, dq subsystem with d and q variables is obtained (x
d
x
q
and u
d
u
q
). The
system equation in (10) becomes as in (11) (Mao et al., 1998; Mihailovic, 1998).
_
X
d
X
q
_ = j
A
I
I
A
[ j
x
d
x
q
[ + j
B
u
u
B
[ j
u
d
u
q
[ (11)
Where I is the identity matrix and 0 is a zero matrix, both having the same dimension as x.
(11) can transform any threephase system into the dq model directly.
Application of InputOutput Linearization 5
When equations (1) (4) are transformed into dq coordinates, (12) (14) are obtained, as in
(Blasko & Kaura, 1997; Ye, 2000; Kazmierkowski et al., 2002).
Ji
d
Jt
= 
R
I
i
d
+ i
q

1
I
I
dc
J
d

u
d
I
(12)
Ji
q
Jt
= 
R
I
i
q
i
d

1
I
I
dc
J
q

u
q
I
(13)
JI
dc
Jt
=
1
C
(i
d
J
d
+i
q
J
q
) 
1
C
i
dc
(14)
Where i
d
and i
q
are the dq transformation of i
u
, i
b
and i
c
. :
d
and :
q
are the dq
transformation of :
u
, :
b
and :
c
. J
d
and J
q
are the dq transformation of J
u
, J
b
and J
c
.
4. Inputoutput feedback linearization technique
Feedback linearization can be used as a nonlinear design methodology. The basic idea is
first to transform a nonlinear system into a (fully or partially) linear system, and then to use
the wellknown and powerful linear design techniques to complete the control design. It is
completely different from conventional linearization. In feedback linearization, instead of
linear approximations of the dynamics, the process is carried out by exact state
transformation and feedback. Besides, it is thought that the original system is transformed
into an equivalent simpler form. Furthermore, there are two feedback linearization methods
that are inputstate and inputoutput feedback linearization (Slotine & Li, 1991; Isidori, 1995;
Khalil, 2000; Lee et al., 2000; Lee, 2003).
The inputoutput feedback linearization technique is summarized by three rules;
Deriving output until input appears
Choosing a new control variable which provides to reduce the tracking error and to
eliminate the nonlinearity
Studying stability of the internal dynamics which are the part of system dynamics
cannot be observed in inputoutput linearization (Slotine & Li, 1991)
If it is considered an inputoutput system, as in (15)(16);
X
i
1
b
(x) = u (24)
y
i
= I
]
i
b
+ [I
g
i
I
]
i
1
b
n
=1
(25)
Applying (25) for all n outputs, (26) is derived.
_
y
1
y
1
n
_ = _
I
]
1
b
1
(x)
I
]
n
b
n
(x)
_ +E(x) _
u
1
u
n
_ = o(x) + E(x)u (26)
E(x) in (27) is a decoupling matrix, if it is invertible and new control variable is chosen,
feedback transformation is obtained, as in (28).
E(x) = _
I
g
1
I
]
1
1
b
1
I
g
n
I
]
n
1
b
1
I
g
1
I
]
n
1
I
g
n
I
]
n
1
b
n
_ (27)
_
u
1
u
n
_ = E
1
_
I
]
1
b
1
(x)
I
]
n
b
n
(x)
_ +E
1
_
:
1
:
n
_ (28)
Equation (29) shows the relation between the new inputs v and the outputs y. The input
output relation is decoupled and linear (Lee et al., 2000).
Application of InputOutput Linearization 7
_
y
1
y
1
n
_ = _
:
1
:
n
_ (29)
If the closed loop error dynamics is considered, as in (30) (31), (32) defines new inputs for
tracking control.
_
c
1
+ k
1(2)
c
1
1
++ k
11
c
1
1
+ k
10
c
1
c
n
+k
n(1)
c
n
1
+ +k
21
c
1
1
+ k
20
c
n
_ = _
u
u
_ (30)
_
c
_ = _
y  y
 y
r
_ (31)
_
:
1
:
n
_ = _
k
1(1)
y
1
  k
11(1)
y
1
 k
10
(y
1
y
1
k
n(1)
y
1
  k
21(1)
y
1
 k
20
(y
n
y
n
)
_ (32)
k values in equations show the constant values for stability of systems and tracking of y
references, as in (Lee, 2003).
5. The application of an inputoutput feedback linearization to the VSC
The state feedback transformation allows the linear and independent control of the d and q
components of the line currents in VSC by means of the new inputs u
d
and u
q
.
For unity power factor, in (12 14) u
d
= I
m
and u
q
= u are taken, so mathematical model of
this system is derived with (3335), as in (Kmrcgil & Kkrer, 1998; Lee, 2003).
Ji
d
Jt
= 
R
I
i
d
+ i
q

1
I
I
dc
J
d

I
m
I
(33)
Ji
q
Jt
= 
R
I
i
q
i
d

1
I
I
dc
J
q
(34)
JI
dc
Jt
=
1
C
(i
d
J
d
+i
q
J
q
) 
1
C
i
dc
(35)
If (3335) are written with the form of (15), (36) is derived.
(x) =
l
l
l
l
l

R
L
i
d
+i
q
+
v
m
L
i
d

R
L
i
q
1
C
i
dc
1
1
1
1
1
, g(x) =
l
l
l
l
l

1
L
I
dc
u
u 
1
L
I
dc
1
C
i
d
1
C
i
q
1
1
1
1
1
(36)
The main purpose of this control method is to regulate I
dc
voltage by setting i
d
current and
to provide unity power factor by controlling i
q
current. Therefore, variables of y outputs
and reference values are chosen as in (37).
Applications of Nonlinear Control 8
y = j
y
1
y
2
[ = _
b
1
(x)
b
2
(x)
_ = _
i
d
i
q
_ , y
= j
I
d
u
[ (37)
Differentiating outputs of (37), (38) is obtained. The order of derivation process, finding a
relation between y outputs and u inputs, is called as relative degree. It is also seen that the
relative degree of the system is 1.
y = _
t
d
t
q
_ = _

1
L
I
dc
u
u 
1
L
I
dc
_ u +_

R
L
i
d
+i
q
+
v
m
L
i
d

R
L
i
q
_ (38)
Fig. 3. Inputoutput feedback linearization control algorithm of VSC
When (38) is ordered like (28), (39) is obtained.
u = j
u
d
u
q
[ = _

1
L
I
dc
u
u 
1
L
I
dc
_
1
. __

R
L
i
d
+i
q
+
v
m
L
i
d

R
L
i
q
_ +:_ (39)
After taking inverse of matrix (39) and adding new control inputs from (40), (41) is obtained.
Application of InputOutput Linearization 9
j
:
1
:
2
[ = _
k
1
(y
1
 I
d
)
k
2
(y
2
I
q
)
_ (40)
j
u
d
u
q
[ = _

L
v
dc
u
u 
L
v
dc
_ . __
R
L
i
d
i
q

v
m
L
i
d
+
R
L
i
q
_ + _
k
1
(y
1
 I
d
)
k
2
y
2
__ (41)
Control algorithm is seen in Fig.3. For both L and LCL filtered VSC, the same control
algorithm can be used. Providing the unity power factor, angle values are obtained from
line voltages. This angle values are used in transformation of abc to dq frames. Line
currents which are transformed into dq frame, are compared with dq reference currents. d
axis reference current i
dc]
is obtained by comparison of dc reference voltage I
c]
and actual
dc voltage I
dc
. On the other hand, q axis reference current i
qc]
is set to 0 to provide unity
power factor. And by using (41), switching functions of dq components are found. Then,
this dq switching functions are transformed into abc frames and they are sent to PWM
block to produce pulses for power switches.
Providing the control of internal dynamics, in dc controller square of I
c]
and I
dc
are used
(Lee, 2003).
6. DC machine and armature circuit model
Electrical machines are used for the conversion of electric power to mechanical power or
vice versa. In industry, there are wide range of electrical machines that are dc machines,
induction machines, synchronous machines, variable reluctance machines and stepping
motors. The Dc machines can be classified as a brushless and brushed dc machines.
Furthermore, the advantage of brushed dc machines is the simplicity with regard to speed
control in the whole machines. However, the main disadvantage of this kind of machines is
the need of maintenance because of its brushes and commutators.
Fig. 4 shows the basic structure of brushed dc machines. Basic components of dc machines
are field poles, armature, commutator and brushes (Fitzgerald et al., 2003).
Fig. 4. Dc machine
Field poles produce the main magnetic flux inside of the machines with the help of the field
coils which are wound around the poles and carry the field current. Some of the dc
Applications of Nonlinear Control 10
machines, the magnetic flux is provided by the permanent magnet instead of the field coils.
In Fig. 5, the field coils and field poles of dc machines are shown (Bal, 2008).
Fig. 5. Field coils and field poles of a dc machine
The rotating part of the dc machine is called as an armature. The armature consists of iron
core, conductors and commutator. Besides, there is a shaft inside of armature that rotates
between the field poles. The other part of the machine is commutator which is made up of
copper segments and it is mounted on the shaft. Furthermore, the armature conductors are
connected on the commutator. Another component of dc machine is brushes. The brushes
provide the electric current required by armature conductors. In dc machine to ensure the
rotation of the shaft, the armature conductors must be energized. This task is achieved by
brushes that contact copper segments of commutator. Also, the brushes generally consist of
carbon due to its good characteristic of electrical permeability. Fig. 6 shows the armature,
commutator and the brushes (Fitzgerald et al., 2003; Bal, 2008).
Fig. 6. a) armature, b) commutator and c) brushes of a dc machine
Application of InputOutput Linearization 11
To produce the main flux, the field must be excited. For this task, there are four methods
which are separately, shunt, series and compound to excitation of dc machines and are
shown in Fig. 7. However, separately excited dc machine is the most useful method because
it provides independent control of field and armature currents. Therefore, this structure is
used in this chapter (Krishnan, 2001; Fitzgerald et al., 2003).
Fig. 7. Excitation methods of dc machine a) separately, b) shunt, c) series, d) compound
excitation
There are two basic speed control structure of dc machine which are armature and field, as
in (Krishnan, 2001). The armature circuit model of dc machine is shown in Fig. 8.
Fig. 8. Armature circuit model of dc machine
Applications of Nonlinear Control 12
The mathematical model of armature circuit can be written by (42).
: = c +R
u
I
u
+I
u
Ji
u
Jt
(42)
In steady state,
d
c
dt
part is zero because of the armature current is constant. The armature
model is then obtained by (43), (Krishnan, 2001).
: = c + R
u
I
u
(43)
c = K1
]
m
(44)
(43) is written in (44), (45) is derived.
m
=
(I  R
u
I
u
)
I
]
(45)
The speed of dc machine depends on armature voltage and field current, as shown in (45). In
field control, the armature voltage is kept constant and the field current is set. The relation
between speed and field current is indirect proportion. However, in armature control, the
relation between armature voltage and speed is directly proportional. Furthermore, in
armature control, the field current is kept constant and the armature voltage is set.
In this chapter, the armature control of dc machine is realized.
The speed control loop is added to nonlinear control loop. Firstly, the actual speed is
compared with reference speed then the speed error is regulated by PI controller and after
that its subtraction from armature current, the reference current is obtained. The reference
current obtained by speed loop, is added to nonlinear control loop instead of reference i
d
current, which is obtained by the comparison of the square of reference voltage and actual
voltage.
In Fig. 9, speed control loop is shown.
Fig. 9. Armature speed control loop of dc machine
7. Simulations
Simulations are realized with Matlab/Simulink. Line voltage is taken 220 V, 60 Hz. The
switching frequency is also chosen 9 kHz. L filter and controllers parameters are shown in
Table.1 and Table.2.
Simulation diagram is shown in Fig.10. By simulation, steadystate error and settling time of
dc motor speed, harmonic distortions and shapes of line currents and unity power factor are
examined.
Application of InputOutput Linearization 13
Passive Components
L Filter DcLink
L (H) R () C
dc
(F)
0.0045 5.5 2200
Table 1. Values of L filter components
Controllers
Speed Controller Inputoutput current controller
K
p
K
k (1u
3
) K (1u
5
)
10 0.01 30 50
Table 2. Values of controllers
Fig. 10. Simulation diagram of dc machine controller in Simulink
Fig.11 shows the structure of inputoutput controller diagram.
Fig. 11. Inputoutput controller diagram
nonl i near control l er
v3 v2 v1
m
speedref
current ref
I armature
speed control l er
Discrete,
Ts = 5e006 s.
powergui
In1
In2
v ia
v ib
v ic
gri d
iq
vdc
id
vabc1
iabc
Pulses
v
+

g
A
B
C
+

Uni versal Bri dge
Tork
Speedref
I
_
a
b
c
v
a
b
c
1
v
a
b
c
2
T
a
T
b
T
c
M
t
a
M
t
b
M
t
c
I_abc
Ta
Tb
Tc
Mta
Mtb
Mtc
Measures
La
Lb
Lc
Lla
Llb
Llc
Inductors
TL m
A+
F+
A
F
dc
0
V bus (V)
abc  dq
1
Pulses
Id
Iq
Idref
Iqref
vdc
ud
uq
input  output
controller
dq0
sin_cos
abc
dq0  abc abc sincos
angle
generator
abc
sin_cos
dq0
Uref Pulses
0
5
iabc
4
vabc1
3
vdc
2
id
1
iq
Applications of Nonlinear Control 14
Equation (41) is written in the block of inputoutput controller which is shown in Fig.12.
Fig. 12. Input output controller
Fig. 13 shows the speed controller of dc machine.
Fig. 13. Speed controller of dc machine
Fig.14 shows the dc machine speed. Reference speed value is changed from 150 rad/s to
200rad/s at 0.5 s. Settling time to the first reference is shorter than 0.15 s, but settling time of
second reference is 0.1 s.
Fig. 15 shows the steadystate error of dc machine speed. It is seen that the steady state
error changes between 2 rad/s.
The one phase voltage and current is shown in Fig. 16. It is also seen that unity power factor
is obtained but not as desired.
Fig. 17 shows the line currents. The shapes of line currents are sinusoidal.
2
uq
1
ud
((L)/u[5])*((w*u[2])+((R*u[1])/L(E/L))(k*(u[1]u[3])))
((L)/u[5])*((w*u[1])+((R*u[2])/L)(K*u[2]))
5
vdc
4
Iqref
3
Idref
2
Iq
1
Id
2
I armature
1
current ref
fi el d
tork
PI
2
speedref
1
m
Application of InputOutput Linearization 15
Fig. 14. Dc machine speed
Fig. 15. Steadystate error of dc machine speed
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
50
100
150
200
250
time(s)
s
p
e
e
d
(
r
a
d
/
s
)
dc machine speed
0.3 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38 0.39 0.4
144
146
148
150
152
154
156
time(s)
s
p
e
e
d
(
r
a
d
/
s
)
dc machine speed
Applications of Nonlinear Control 16
Fig. 16. One phase voltage and current
Fig. 17. Threephase phase current
0.28 0.29 0.3 0.31 0.32 0.33 0.34 0.35
300
200
100
0
100
200
300
time(s)
M
a
g
n
i
t
u
d
e
,
(
V
)
,
(
A
)
voltage
current
0.13 0.135 0.14 0.145 0.15
40
30
20
10
0
10
20
30
40
time(s)
C
u
r
r
e
n
t
(
A
)
phase a
phase b
phase c
Application of InputOutput Linearization 17
Fig. 18 shows the harmonic distortions of line currents. Line currents include high order
harmonic contents. However, total harmonic distortion value (THD) is under the value that
is defined by standards. THD of line currents are %1.34, %1.71 and %2.84.
Fig. 18. Harmonic distortions of line currents
I
a
I
b
Applications of Nonlinear Control 18
8. Conclusion
In this chapter, simulation of dc machine armature speed control is realized. Dc machine is
fed by voltage source rectifier which is controlled input output linearization nonlinear
control method. Furthermore, for the speed control, dc link voltage is regulated by the dc
machine speed control loop. The control algorithm of voltage source rectifier and dc motor
speed are combined. The required reference I
d
current for voltage source rectifier is obtained
by speed control loop. Simulations are carried through Matlab/Simulink. By means of the
simulation results, the speed of dc machine, line currents harmonic distortions and power
factor of grid are shown. It is shown that the voltage source rectifier with dc machine as a
load provides lower harmonic distortion and higher power factor. Furthermore, dc machine
speed can be regulated.
9. References
Bal, G. (2008). Dogru Akm Makinalar ve Suruculeri, Seckin, ISBN 9789750207068, Sihhiye,
Ankara, Turkey
Bates, J.; Elbuluk, M,E. & Zinger, D,S. (1993). Neural Network Control of a ChopperFed Dc
Motor, IEEE Power Electronics Specialist Conference PESC02, Vol 3, pp. 893, ISBN 0
780312430, Seattle, Washington, USA, June, 1993
Blasko, V. & Kaura, V. (1997). A New Mathematical Model and Control ThreePhase acdc
Voltage Source Converter, IEEE Transactions on Power Electronics, Vol 12, January,
1997, pp. 7881, ISSN 08858993
Bose, B,K. (2002). Modern Power Electronics and Ac Drives, Prentice Hall, ISBN 0130167436,
New Jersey, USA
Dai, K.; Liu, P. ; Kang, Y. & Chen, J. (2001). Decoupling Current Control for Voltage Source
Converter in Synchronous Rotating Frame, IEEE PEDS, pp. 3943, ISBN 07803
72336, Indonesia, JanuaryFebruary, 2001
Dannehl, J.; Fuchs, F,W. & Hansen, S. (2007). PWM Rectifier with LCL filter Using Different
Control Structures, EPE Aalborg, pp. 110, ISBN 9729275815102, Aalborg,
October, 2007
Dixon, J,W. & Ooi, B,T. (1988). Indirect Current Control of a Unity Power Factor Sinusoidal
Current Boost Type Threephase Rectifier, IEEE Transaction of Indisturial Electronics,
Vol.35, Nov, 1988, pp. 508515, ISSN 02780046
Dixon, J,W. (1990). Feedback Control Strategies for Boost Type PWM Rectifiers . Proceedings
of IEEE Colloquium in South America, pp. 193198, ISBN 0879426101, September,
1990
Fitzgerald, A,E.; Kingsley, C,Jr. & Umans, S,D. (2003). Electric Machinery, McGraw Hill, ISBN
0071121935, New York, USA
Holtz, J. (1994). Pulse Width Modulation for Electronic Power Conversion, Proceedings of the
IEEE, Vol 82, August 1994, pp. 11941214, ISSN 00189219
Isidori, A. (1995). Nonlinear Control Systems, SpringerVerlag, ISBN 3540199160,
Heidelberg New York
Kazmierkowski, M,P.; Krishnan, R. & Blaabjerg, F. (2002). Control in Power Electronics:
Selected Problems, Elsevier Science, ISBN 0124027725, San Diego, California,
USA
Application of InputOutput Linearization 19
Kim, D,E. & Lee, D,C. (2007). Feedback Linearization Control of ThreePhase Ac/Dc PWM
Converters with LCL Input Filters, International Conference on Power Electronics
ICPE07, pp. 766771, ISBN 9781424418718, Daegu, South Korea, October,
2007
Khalil, H,K. (2000). Nonlinear Systems, Pearson Education, ISBN 0131227408, New Jersey,
USA
Krishnan, R. (2001). Electric Motor Drives: Modeling, Analysis, and Control, Prentice Hall, ISBN
0130910147, New Jersey, USA
Kmrcgil, H. & Kkrer, O. (1998). Lyapunov Based Control of ThreePhase PWM AcDc
VoltageSource Converters, IEEE Transactions on Power Electronics, Vol 13,
September 1998, pp. 801813, ISSN 08858993
Lee, D,C.; Lee, G,M. & Lee, K,D. (2000). Dc Bus Voltage Control of Three Phase AcDc PWM
Converters Using feedback Linearization, IEEE Transactions on Industry Application,
Vol 36, MayJune 2000, pp. 826833, ISSN 09939994
Lee, T,S. (2003). InputOutput Linearizing and Zero Dynamics Control of ThreePhase Ac
Dc Voltage Source Converters, IEEE Transactions on Power Electronics, Vol 18,
January 2003, pp. 1122, ISSN 08858993
Lindgren, M. (1998). Modelling and Control of Voltage Source Converters Connected to the
Grid, PHDs Thesis Faculty of Chalmers University of Technology, ISBN 9171977104,
Goteborg, SWEEDEN, November, 1998
Liserre, M.; Blaabjerg, F. & Hansen, S. (2005). Design and Control of an LCL Filter Based
ThreePhase Active Rectifier, IEEE Transactions on Industry Application, Vol 41,
SeptemberOctober 2005, pp. 12811291, ISSN 09939994
Mao, H.; Boroyevich, D. & Lee, F.C. (1998). Novel ReducedOrder SmallSignal Model of
ThreePhase PWM Rectifiers and its Application in Control Design and System
Analysis, IEEE Transactions on Power Electronics, Vol 13, May 1998, pp. 511521,
ISSN 08858993
Mihailovic, Z. (1998). Modeling and Control Design of VSI_ Fed PMSM Drive Systems with
Active Load, Master Thesis Faculty of Virginia Polytechnic and State University,
Blacksburg, Virginia, USA, September, 1998
Ooi, B,T.; Salmon, J,C. ; Dixon, J,W. & Kulkarini, A,B. (1987). A threephase Controlled
Current PWM Converter with Leading Power Factor, IEEE Transaction of Indistury
Application, Vol.IA23, JanFeb 1987, pp. 7881, ISSN 00939994
Sehirli, E. & Altnay, M. (2010). Simulation of ThreePhase Voltage Source Pulse Width
Modulated (PWM) LCL Filtered Rectifier Based on InputOutput Linearization
Nonlinear Control, IEEE International Conference on Electrical and Electronic
Equipment OPTIM2010, pp. 564569, ISBN 9781424470198, Brasov, Romania,
May, 2010
Slotine, J,J,E. & Li, W. (1991). Applied Nonlinear Control, Prentice Hall, ISBN 0130408905,
New Jersey, USA
Sousa, G,C,D. & Bose, B. (1994). A Fuzzy Set Theory Based Control of a Phase  Controlled
Converter Dc Machine Drive, IEEE Transactions on Industry Application, Vol 30,
JanuaryFebruary, 1994, pp. 3444, ISSN 09939994
Wu, R.; Dewan, S,B. & Slemon, G,R. (1988). A PWM ac to dc Converter with Fixed Switching
Frequency, Conference Recordings 1988 IEEEIAS Annual Meeting, Vol 1, October,
1988 pp. 706711, Pittsburgh, PA, USA
Applications of Nonlinear Control 20
Wu, R.; Dewan, S,B. & Slemon, G,R. (1991). Analysing of ac to dc Voltage Source
Converter Using PWM with Phase and Amplitude Control, IEEE Transactions
Industirial Applications, Vol 27, MarchApril, 1991, pp. 355364, ISSN 0093
9994
Ye, Z. (2000). Modelling and Control of Parallel ThreePhase PWM Converters, PHDs Thesis
Faculty of Virginia Polytechnic and State University, pp. 920, Blacksburg, Virginia,
USA, September, 2000
0
LyapunovBased Robust and Nonlinear
Control for TwoStage Power Factor
Correction Converter
Seigo Sasaki
National Defense Academy
Japan
1. Introduction
Many power electronic system designs focus on energy conversion circuit parameters rather
than controller parameters which drive the circuits. Controllers must be designed on the
basis of circuit models, which are generally nonlinear systems (Brockett & Wood (1974)), in
order to improve performance of controlled systems. The performance of controlled systems
depend on nominal models to design the controllers. More broad class of models controllers
are designed for, better control performance may be given. Many works (e.g. Kassakian
et al. (1991)) design controllers for linearized models because it is not easy to concretely
design controllers for the nonlinear models. Controller design in consideration of nonlinear
models has been discussed since a work by Banerjee & Verghese (2001) because a research on
nonlinear controller design has grown in those times.
This chapter systematically designs a robust controlled power converter system on the
basis of its nonlinear model. Concretely, a twostage power factor correction converter,
that is a forward converter (FC) with power factor corrector (PFC), is designed. The
systematic controller design clearly analyzes the behavior of nonlinear system to improve
the performance.
A work by Orabi & Ninomiya (2003) analyzes a stability of singlestage PFC for variations
of controller gain on the basis of its nonlinear model. On the basis of the work by Orabi &
Ninomiya (2003) that regards a load of PFC as constant, a work by Dranga et al. (2005) for a
twostage PFC focuses on a point that a load of PFC part is not pure resistive and analyzes a
stability of the converter.
A work by Sasaki (2002) discusses an inuence between a FC part and a PFC part in
a twostage PFC. A work by Sasaki (2009) clearly shows that a source current reference
generator plays an important role in a synthesis for a singlestage PFC. For the works by
Sasaki (2002; 2009), this chapter shows how to decide synthesis parameters of robust linear
and nonlinear controllers for a twostage PFC in more detail.
The controller synthesis step, that is this chapter, is organized as follows. First, the converter is
divided into two parts which consists of a FCpart and a PFCpart by considering an equivalent
2
2 Nonlinear Control
circuit of transformer. The two parts depend on each other and are nonlinear systems. The
FC part has an apparent input voltage which depends on an output voltage in the PFC part.
On the other hand, the PFC part has an apparent load resistance which depends on an input
current in the FC part. Second, the two parts of converter are treated as two independent
converters by analyzing steady state in the converter and deciding a set point. Then, the
above input voltage and load resistance are xed on the set point. Controllers are designed
for the two independent converters respectively. For the FC part which is a linear system at
the second stage of converter, a robust linear controller is designed against variations of the
apparent input voltage and a load resistance. For the PFC part which is a bilinear system,
that is a class of nonlinear system, at the rst stage, a robust nonlinear controller is designed
against variations of the apparent load resistance that mean dynamic variations of the FC part
at the second stage. Finally, computer simulations demonstrate efciencies of the approach.
It is also claried that consideration of nominal load resistance for each part characterizes a
performance of the designed robust controlled system.
2. Twostage power factor correction converter
A controlled system for a twostage power factor correction converter, that is a forward
converter (FC) with power factor corrector (PFC), is systematically constructed as shown in
Fig.1. The systematic controller design clearly analyzes the behavior of controlled converter
system. In this section, rst, an equivalent circuit of transformer divides the converter into two
parts with FC and PFC. Averaged models for the two parts are derived respectively, which
depend on each other. Second, steady state in the two parts is analyzed, which is used as a set
point. Finally, each of parts is treated as an independent converter respectively.
v
s
v
2
i
2
v
1
i
1
L
2
r
2
C
2
L
1
r
1
C
1
1 : N
R
S
2
S
1
2
1
Fig. 1. Twostage converter ; forward converter (FC) with power factor corrector (PFC)
2.1 Nonlinear averaged models
An equivalent circuit of transformer derives a circuit as shown in Fig. 2 fromFig. 1. A FC part
has a variable dc source voltage which depends on a switch S
2
. APFC part has a variable load
resistance which depends on a switch S
1
. Fig. 2 gives two nonlinear averaged models (
f c
SA
)
and (
p f c
SA
) for the FC and the PFC parts respectively, which depend on each other.
The FC model (
f c
SA
) is given as
d
dt
_
v
1
i
1
_
=
_
1
RC
1
1
C
1
1
L
1
r
1
L
1
_
_
v
1
i
1
_
+
_
v
2
_
0
N
L
1
__
1
(1)
22 Applications of Nonlinear Control
LyapunovBased Robust and Nonlinear Control for TwoStage Power Factor Correction Converter 3
v
s
v
2
i
2
v
1
i
1
L
2
r
2
C
2
L
1
r
1
C
1 R
S
2
S
1
S
1
Ni
1
Nv
2
2
1
R
2
(
1
)
E(
2
)
PFC FC
Fig. 2. Two converters ; FC and PFC
and the PFC model (
p f c
SA
) is
d
dt
_
v
2
i
2
_
=
_
0
1
C
2
1
L
2
r
2
L
2
_
_
v
2
i
2
_
+
_
0
1
L
2
_
v
s

+
_
v
2
_
0
1
L
2
_
+
i
2
_
1
C
2
0
__
2
+
_
i
1
_
N
C
2
0
__
1
(2)
where v
1
and v
2
are averaged capacitor voltages,
i
1
,
i
2
averaged inductor currents, C
1
,C
2
capacitances, L
1
,L
2
inductances, r
1
,r
2
internal resistances, R a load resistance, N a turns ratio,
v
s
a source voltage,
1
,
2
averaged switching functions given by 0 <
1
< 1 and 0 <
2
< 1.
2.2 DC components of steady state
DC components of steady state in two parts of converter are derived, which are used as a
set point for controller design. Given an average fullwave rectied voltage V
s
:= (/2)
_
2
0
v
s
(t)dt = 2
2V
e
/ of a source voltage v
s
=
2 V
e
sint, and specify averaged
switching functions
1
=
1s
,
2
=
2s
which are called a set point, then dc components
of voltages and currents of steady state are given by
v
1s
:=
1s
1
1 +
r
1
R
E(
2s
), (3)
i
1s
:=
1
R
v
1s
=
1s
1
1 +
r
1
R
E(
2s
)
R
, (4)
v
2s
:=
1
1
2s
+
1
1
2s
r
2
R
2
(
1s
)
V
s
, (5)
i
2s
:=
1
1
2s
1s
N
i
1s
=
1
1
2s
v
2s
R
2
(
1s
)
(6)
where
E(
2s
) := N v
2s
, R
2
(
1s
) :=
1 +
r
1
R
2
1s
N
2
R. (7)
23
LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter
4 Nonlinear Control
The above steady states in two parts depend on each other through (7). It is easily claried by
assuming r
1
= r
2
= 0 that the equations (3), (4) give a steady state in buck converter and the
equations (5), (6) give a steady state in boost converter.
2.3 Two independent models
For two parts of converter with FCand PFC, the steady state analysis derives two independent
converters. It means that the FC has an apparent source voltage E(
2s
) and the PFC has an
apparent load resistance R
2
(
1s
) which are xed on a set point respectively.
Then, two independent averaged models of the converters are given as
a linear FC model (
f c
A0
) of the form
d
dt
v
1
i
1
1
RC
1
1
C
1
1
L
1
r
1
L
1
_
v
1
i
1
_
+
0
E(
2s
)
L
1
1
(8)
and a nonlinear PFC model (
p f c
A0
) of the form
d
dt
v
2
i
2
1
R
2
(
1s
)C
2
1
C
2
1
L
2
r
2
L
2
v
2
i
2
0
1
L
2
v
s
 +
v
2
0
1
L
2
+
i
2
1
C
2
0
2
. (9)
The FC is modeled on a linear system and the PFC is modeled on a bilinear system that is a
class of nonlinear system (Brockett & Wood (1974); Mohler (1991)).
Here derived is an amplitude of source current of steady state in the PFC, which is used for
source current reference generator in Section 3.3. Given a source voltage v
s
=
2V
e
sin t [V],
assume an output voltage v
2
= v
2r
[V], then dc components of steady state in the PFC gives a
source current
i
2
=
2I
e
sint [A], which is sinusoidal and in phase with the source voltage,
given by
I
e
=
V
e
2r
2
V
2
e
4r
2
2
v
2
2r
r
2
R
2
(
1s
)
. (10)
This decision process is similar to that is discussed in works by Escobar et al. (1999); Escobar
et al. (2001).
3. Robust controller design
Controllers for two parts with FC and PFC are designed on the basis of two independent
converter models (
f c
A0
) and (
p f c
A0
) respectively. First, the models (
f c
A0
) and (
p f c
A0
) are moved
to set points given in Section 2.2. Next, controller design models, that are called as generalized
plants, are derived which incorporate weighting functions in consideration of controller
design specications. Finally, feedback gains are given to guarantee closedloop stability and
reference tracking performance.
24 Applications of Nonlinear Control
LyapunovBased Robust and Nonlinear Control for TwoStage Power Factor Correction Converter 5
3.1 Models around set points
Converter models around set points are derived from the averaged models (
f c
A0
) and (
p f c
A0
).
Moving the state to a specied set point given by v
1
:= v
1
v
1s
,
i
1
:=
i
1
i
1s
,
1
:=
1
1s
,
v
2
:= v
2
v
2s
,
i
2
:=
i
2
i
2s
,
2
:=
2
2s
derives averaged models around the set point
respectively for the two parts, which are given by
a FC model (
f c
A
) of the form
d
dt
_
v
1
i
1
_
=
1
RC
1
1
C
1
1
L
1
r
1
L
1
_
v
1
i
1
_
+
0
E(
2s
)
L
1
1
(11)
=: A
f c
p
x
f c
p
+ B
f c
p
u
1
(12)
and a PFC model (
p f c
A
) of the form
d
dt
_
v
2
i
2
_
=
1
R
2
(
1s
)C
2
1
C
2
(1
2s
)
1
L
2
(1
2s
)
r
2
L
2
_
v
2
i
2
_
+
_
0
1
L
2
_
(v
s
 Vs)
+
__
i
2s
C
2
v
2s
L
2
_
+ v
2
_
0
1
L
2
_
+
i
2
_
1
C
2
0
__
2
(13)
=: A
p f c
p
x
p f c
p
+ B
p f c
p1
w
p f c
p
+
_
B
p f c
p
+
_
x
p f c
p
N
p f c
p
__
u
2
(14)
where
1s
<
1
< 1
1s
and
2s
<
2
< 1
2s
. The following discussion constructs
controllers on the basis of the averaged models (
f c
A
) and (
p f c
A
) around the set point.
3.2 Linear controller design for FC
A robust controller for the FC is given by linear H
T
and x
f c
w
denotes state
of weighting function W
v1
(s) as shown in Fig.3. In the gure, matrices W
e1
and W
u1
are
weighting coefcients of a performance index in the linear H
[ 1 0 ]
Linear
Gain
controller
Fig. 3. Block diagram of linear controlled system for FC
_
A
f c
Y
f c
+ Y
f c
A
f c
T
B
f c
2
(D
f c
12
T
D
f c
12
)
1
B
f c
2
T
+
f c
2
B
f c
1
B
f c
1
T
Y
f c
C
f c
1
T
C
f c
1
Y
f c
I
_
< 0 (16)
where the matrix A
f c
,B
f c
1
, C
f c
1
are coefcients of a generalized plant given by Fig.3 and
f c
is
a synthesis parameter in the linear H
control technique.
3.3 Nonlinear controller design for PFC
A robust nonlinear controller design for the PFC, that is a nonlinear system, consists of the
following two steps. First, a nonlinear gain is given to guarantee closed loop system stability
and reference tracking performance for source current and output voltage against variations of
the apparent load resistance. Second, a source current reference generator is derived to adjust
an amplitude of current reference to variations of source voltage and load resistance. At the
rst step, as shown in Fig. 4 incorporated are a weighting function W
i2
(s) := k
i2
2
i2
/(s
2
+
2
i2
s +
2
i2
) for a source current to be sinusoidal and be in phase with a source voltage and
a function W
v2
(s) := k
v2
/(s +
2
) for an output voltage to be kept constant against those
variations, where k
i2
, ,
i2
, k
v2
,
2
are synthesis parameters. Then, a nonlinear H
control
technique in a work by Sasaki & Uchida (1998) gives a nonlinear gain of the form
K
2
(x
p f c
p
) := (D
p f c
12
T
D
p f c
12
)
1
B
p f c
2
(x
p f c
)
T
Y
p f c
1
(17)
as shown in Fig.4, where x
p f c
=
_
x
p f c
p
T
x
p f c
w
T
_
T
, D
p f c
12
=
_
0
T
W
T
u2
T
, B
p f c
2
(x
p f c
) =
_
B
p f c
p
0
_
+
x
p f c
1
_
B
p f c
p21
0
_
+ x
p f c
2
_
B
p f c
p22
0
_
, B
p f c
p21
=
_
0
1
L
2
_
T
, B
p f c
p22
=
_
1
C
2
0
_
T
and x
p f c
w
denotes state of
weighting functions. In the gure, matrices W
e2
and W
u2
are weighting coefcients of a
performance index in the nonlinear H
i
2
v
2
x
p f c
w2,3
x
p f c
w1
u
2
i
2r
v
2r
z
e2
z
u2
w
p f c
p
W
e2
W
u2
B
p f c
p1
W
i2
(s)
W
v2
(s)
I
e
(v
2r
)
1/
2V
e
2
K
2
(x
p f c
p
)
1
s
K(s)
v
s
+
+
+
+
+
+
+
+
+
Nonlinear
Gain
Feedforward
Feedback
Generator
Controller
Fig. 4. Block diagram of nonlinear controlled system for PFC
The matrix Y
p f c
is given as a positivedenite solution satisfying a statedepended
Lyapunovbased inequality condition
A
p f c
Y
p f c
+ Y
p f c
A
p f c
T
B
p f c
2
(x)(D
p f c
12
T
D
p f c
12
)
1
B
p f c
2
(x)
T
Y
p f c
C
p f c
1
T
+
p f c
2
B
p f c
1
B
p f c
1
T
C
p f c
1
Y
p f c
I
< 0 (18)
where x is a state of a generalized plant given by Fig.4, matrices A
p f c
,B
p f c
1
,B
p f c
2
(x),C
p f c
1
,D
p f c
12
are coefcients of the plant and
p f c
is a synthesis parameter in the nonlinear H
control
technique.
For any state x, which is current and voltage in a specied domain, the matrix Y
p f c
satisfying
the inequality (18) is concretely given by solving linear matrix inequalities at vertices of a
convex hull enclosing the domain as shown in a work by Sasaki & Uchida (1998).
Next, given is a mechanism to generate a source current reference i
2r
. As shown in Fig.4
the reference generator consists of a feedforward loop given by steady state analysis in
Section 2.3 and a feedback loop with a voltage error amplier. The amplier K(s) is given
by K(s) = k
P
+ k
I
/s + k
D
s where k
P
,k
I
and k
D
are constant parameters decided by system
designers as discussed in a work by Sasaki (2009). The feedback loop is the same structure
as a conventional loop used in many works (e.g., Redl (1994)). Note that the amplier K(s)
works only for variations fromthe nominal values in the circuit, because the feedforward loop
gives the effective value of the source current as shown in Section 4.4.
4. Computer simulations
This section nally shows efciencies of the approach through computer simulations. It
is also claried that consideration of nominal load resistance for each part characterizes a
performance of the designedcontrolled system. Asoftware package that consists of MATLAB,
Simulink and LMI Control Toolbox is used for the simulations.
27
LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter
8 Nonlinear Control
Parameters of the circuit shown in Fig.1 are given by Table 1.
r
1
1 [m] r
2
1 [m]
L
1
10 [mH] L
2
1 [mH]
C
1
450 [F] C
2
1000 [F]
N 1/36
Table 1. Circuit parameters of twostage power factor correction converter as shown in Fig.1
4.1 Design specication
Control system design specication for the twostage power factor correction converter is
given by
(1) An effective value of source voltage is 82 255 volts, and its frequency is 45 65 hertz ;
(2) An output voltage is kept be 5 volts, and its error of steady state is within 0.5%;
(3) An output current is 0 30 amperes ;
(4) A source current is approximately sinusoidal and is phase with source voltage.
The above specications are treated for controller design as the following respective
considerations ;
(1) A nonlinear gain for PFC is designed for a source voltage whose nominal effective value
is V
e
= (82 +255)/2 = 168.5 volts (then, average fullwave rectied voltage is V
s
= 151.7
volts). Moreover, the gain is designed to be robust against source voltage variations by
treating that as a disturbance w
p f c
p
as shown in Fig.4.
(2) Incorporated is a weighting function W
v1
(s) whose magnitude is high at low frequencies.
(3) A load resistances R varies between 1/6 ohms at an output voltage of 5 volts.
(4) Incorporated is a weighting function W
i2
(s) whose magnitude is high at frequencies in
the range of 45 65 hertz for the source current to be approximately sinusoidal at the
frequencies.
4.2 Nominal load resistance
Now, a nominal value of load resistance R need be decided to design controllers. The nominal
value inuences a performance of closedloop system controlled by the designed controllers.
Root loci of linearized converters for the two parts as shown in Figs. 5 and 6 , which are
pointwise eigenvalues of the systemmatrices for variations of load resistance R, show that the
larger the load resistance R is, the more oscillatory the behavior of FC is and the slower the
transient response in PFC is. Controllers, here, are constructed in consideration of undesirable
behavior of the converter system. Therefore, a FC controller is designed for a system with
oscillatory behavior, that is for a large load resistance, so that an output voltage does not
oscillate. A PFC controller is designed for a system with fast response, that is for a small load
resistance, so that a source current is not distorted by the controller responding well to source
voltage variations.
28 Applications of Nonlinear Control
LyapunovBased Robust and Nonlinear Control for TwoStage Power Factor Correction Converter 9
14000 12000 10000 8000 6000 4000 2000 0
500
400
300
200
100
0
100
200
300
400
500
Re
I
m
R = 1/6 R = 1/6
R = 500
R = 500
Oscillatory
Asymptotic
Fig. 5. Root locus of coefcient A
f c
p
of linear model for FC
1 0.8 0.6 0.4 0.2 0
1000
800
600
400
200
0
200
400
600
800
1000
Re
I
m
R = 1/6
R = 1/6
R = 500
R = 500
Fast Slow
Fig. 6. Root locus of coefcient A
p f c
p
of linearized model for PFC
4.3 Linear controller design parameters for FC
A FC controller is designed for a large nominal load resistance R = 100 ohms. Consider a set
point
1s
= 0.5 for a capacitor voltage of v
2s
= 360 volts. DC components of steady state are
given by v
1s
= 4.99 volts and
i
1s
= 0.0499 amperes. For a model around the above set point,
incorporated are weighting functions with parameters given by K
v1
= 20,
1
= 0.001 W
e1
= 5,
W
u1
= 1.5 and
f c
= 0.95 where
f c
denotes control performance level given in the linear H
2.00 10
4
1.55 10
3
5.05 10
2
1.55 10
3
7.68 10
2
2.59 10
1
5.05 10
2
2.59 10
1
2.51 10
1
. (19)
29
LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter
10 Nonlinear Control
10
5
10
4
10
3
10
2
10
1
10
0
10
1
20
0
20
40
60
80
100
Frequency [Hz]
M
a
g
n
i
t
u
d
e
[
d
B
]
10
5
10
4
10
3
10
2
10
1
10
0
10
1
100
80
60
40
20
0
Frequency [Hz]
P
h
a
s
e
[
d
e
g
]
Fig. 7. Bode plots of weighting function W
v1
(s) for output voltage tracking in FC controller
design
4.4 Nonlinear controller design parameters for PFC
A PFC controller is designed for a small nominal load resistance R = 10 ohms. Consider a
set point
2s
= (v
2r
V
s
)/v
2r
= 0.579 to make a rectied voltage v
2r
= 360 volts. Then, an
apparent load resistance in the PFC is given by R
2
(
1s
) = 51845.2 ohms for a set point
1s
=
0.5 in the FC. Therefore, dc components of steady state are given by v
2s
= 359.9 volts and
i
2s
=
10
0
10
1
10
2
10
3
10
4
100
50
0
50
Frequency [Hz]
M
a
g
n
i
t
u
d
e
[
d
B
]
10
0
10
1
10
2
10
3
10
4
200
150
100
50
0
Frequency [Hz]
P
h
a
s
e
[
d
e
g
]
Fig. 8. Bode plots of weighting function W
i2
(s) for achieving unity power factor in PFC
controller design
30 Applications of Nonlinear Control
LyapunovBased Robust and Nonlinear Control for TwoStage Power Factor Correction Converter 11
10
4
10
3
10
2
10
1
10
0
10
1
100
80
60
40
20
0
20
Frequency [Hz]
M
a
g
n
i
t
u
d
e
[
d
B
]
10
4
10
3
10
2
10
1
10
0
10
1
150
100
50
0
Frequency [Hz]
P
h
a
s
e
[
d
e
g
]
Fig. 9. Bode plots of weighting function W
v2
(s) for capacitor voltage regulation in PFC
controller design
0.0165 amperes. For a model around the above set point, incorporated are weighting functions
with parameters given by = 0.001,
i2
= 130, K
i2
= 300/
i2
, K
v2
= 0.002,
2
= 0.001,
W
e2
= diag [10
7
, 1], W
u2
= 20 and
p f c
= 0.98 where
p f c
denotes control performance level
given in the nonlinear H
2.77 10
4
5.56 10
4
2.93 10
0
4.70 10
3
1.72 10
4
5.56 10
4
1.89 10
5
8.86 10
1
5.57 10
3
4.03 10
4
2.93 10
0
8.86 10
1
9.71 10
8
2.60 10
2
1.89 10
1
4.70 10
3
5.57 10
3
2.60 10
2
1.10 10
3
2.98 10
3
1.72 10
4
4.03 10
4
1.89 10
1
2.98 10
3
1.24 10
4
(20)
by considering a voltage variation of 10 volts and a current variation of 3 amperes around
the set point (i.e., by considering 349.9 v
2
369.9 and 2.9835 i
2
3.0165) and solving
a convex programming problem as shown in the work by Sasaki & Uchida (1998).
Next, a source current reference generator is constructed with a voltage error amplier given
by k
P
= 8 10
2
, k
I
= 4, k
D
= 1 10
5
as shown in Fig.10. The gain is chosen in order to be
low around twice the input line frequency.
31
LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter
12 Nonlinear Control
10
1
10
0
10
1
10
2
10
3
10
4
10
5
30
20
10
0
10
20
Frequency [Hz]
M
a
g
n
i
t
u
d
e
[
d
B
]
10
1
10
0
10
1
10
2
10
3
10
4
10
5
100
50
0
50
100
Frequency [Hz]
P
h
a
s
e
[
d
e
g
]
Fig. 10. Bode plots of voltage error amplier K(s) in PFC controller
4.5 Simulation results
Figs. 1113 show behaviors of the nonlinear averaged models of FC (
f c
SA
) and PFC (
p f c
SA
) in
the following cases ;
(C1) A load resistance R changes from 1000 to 0.25 ohms in steady state for a source voltage v
s
=
2.92 10
4
1.28 10
3
6.26 10
2
1.28 10
3
8.07 10
2
3.57 10
1
6.26 10
2
3.57 10
1
2.27 10
1
(21)
32 Applications of Nonlinear Control
LyapunovBased Robust and Nonlinear Control for TwoStage Power Factor Correction Converter 13
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
1
2
3
4
5
6
7
time t [s]
v
o
lt
a
g
e
v
1
[
V
]
voltage v
1
reference v
1r
(a) output voltage v
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
5
10
15
20
25
time t [s]
c
u
r
r
e
n
t
i 1
[
A
]
(b) output current i
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time t [s]
F
C
c
o
n
t
r
o
lle
r
o
u
t
p
u
t
u
1
(c) FC controller output u
1
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
355
356
357
358
359
360
361
362
363
time t [s]
v
o
lt
a
g
e
v
2
[
V
]
voltage v
2
reference v
2r
(d) capacitor voltage v
2
0.3 0.35 0.4 0.45 0.5 0.55 0.6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
time t [s]
c
u
r
r
e
n
t
i 2
[
A
]
current i
2
reference i
2r
source voltage  v
s
/ 100
(e) source current i
2
0.3 0.35 0.4 0.45 0.5 0.55 0.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time t [s]
P
F
C
c
o
n
t
r
o
lle
r
o
u
t
p
u
t
u
2
(f) PFC controller output u
2
Fig. 11. (C1) Behaviors when a load resistance R changes from 1000 to 0.25 ohms in steady
state
and
Y
p f c
=
2.76 10
4
5.56 10
4
2.93 10
0
4.70 10
3
1.72 10
4
5.56 10
4
1.89 10
5
8.85 10
1
5.56 10
3
4.03 10
4
2.93 10
0
8.85 10
1
9.71 10
8
2.60 10
2
1.89 10
1
4.70 10
3
5.56 10
3
2.60 10
2
1.10 10
3
2.98 10
3
1.72 10
4
4.03 10
4
1.89 10
1
2.98 10
3
1.24 10
4
. (22)
33
LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter
14 Nonlinear Control
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
1
2
3
4
5
6
7
time t [s]
v
o
lt
a
g
e
v
1
[
V
]
voltage v
1
reference v
1r
(a) output voltage v
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
5
10
15
20
25
time t [s]
c
u
r
r
e
n
t
i 1
[
A
]
(b) output current i
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time t [s]
F
C
c
o
n
t
r
o
lle
r
o
u
t
p
u
t
u
1
(c) FC controller output u
1
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
355
356
357
358
359
360
361
362
363
time t [s]
v
o
lt
a
g
e
v
2
[
V
]
voltage v
2
reference v
2r
(d) capacitor voltage v
2
0.3 0.35 0.4 0.45 0.5 0.55 0.6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
time t [s]
c
u
r
r
e
n
t
i 2
[
A
]
current i
2
reference i
2r
source voltage  v
s
/ 100
(e) source current i
2
0.3 0.35 0.4 0.45 0.5 0.55 0.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time t [s]
P
F
C
c
o
n
t
r
o
lle
r
o
u
t
p
u
t
u
2
(f) PFC controller output u
2
Fig. 12. (C1) Behaviors when a load resistance R changes from 1000 to 0.25 ohms in steady
state where a controller is designed for a different nominal load resistance from that in Fig.11
Entries of the matrix Y
f c
in (21) are clearly different from those in (19). On the other hand,
difference of the Y
p f c
between the matrix (22) and (20) is much small. Those differences give
the different behaviors as mentioned above.
Fig.12 demonstrates the discussion in Section 4.2 that controllers needs be designed in
consideration of undesirable behavior of system. For a large load resistance the output voltage
of the FC oscillates and for small load resistance the source current and the capacitor voltage
of the PFC respond to variations larger than those in Fig.11.
34 Applications of Nonlinear Control
LyapunovBased Robust and Nonlinear Control for TwoStage Power Factor Correction Converter 15
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
1
2
3
4
5
6
7
time t [s]
v
o
lt
a
g
e
v
1
[
V
]
voltage v
1
reference v
1r
(a) output voltage v
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
1
2
3
4
5
6
time t [s]
c
u
r
r
e
n
t
i 1
[
A
]
(b) output current i
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time t [s]
F
C
c
o
n
t
r
o
lle
r
o
u
t
p
u
t
u
1
(c) FC controller output u
1
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
355
356
357
358
359
360
361
362
363
time t [s]
v
o
lt
a
g
e
v
2
[
V
]
voltage v
2
reference v
2r
(d) capacitor voltage v
2
0.3 0.35 0.4 0.45 0.5 0.55 0.6
0
0.1
0.2
0.3
0.4
0.5
time t [s]
c
u
r
r
e
n
t
i 2
[
A
]
current i
2
reference i
2r
source voltage  v
s
/ 280
(e) source current i
2
0.3 0.35 0.4 0.45 0.5 0.55 0.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time t [s]
P
F
C
c
o
n
t
r
o
lle
r
o
u
t
p
u
t
u
2
(f) PFC controller output u
2
Fig. 13. (C2) Behaviors when an efcient value of source voltage V
e
changes from 100 to 85
volts for 2 seconds in steady state
5. Conclusion
For a twostage power factor correction converter, a nonlinear controlled converter system
was systematically designed on the basis of its nonlinear model. The systematic controller
design clearly analyzed the behavior of nonlinear system to improve the performance. The
synthesis step were clearly shown. It was claried that a nominal load resistance characterizes
the controller performance. Finally, computer simulations demonstrated efciencies of
the approach. The nonlinear controller synthesis was shown as a natural extension of a
wellknown Lyapunovbased linear controller synthesis.
35
LyapunovBased Robust and Nonlinear Control
for TwoStage Power Factor Correction Converter
16 Nonlinear Control
6. Acknowledgment
The author would like to thank Hideho Yamamura of Hitachi,Ltd. and Kazuhiko Masamoto
of NEC Networks for their valuable discussions and comments in the viewpoint of industrial
practice.
7. References
Brockett, R.W. & Wood, J.R. (1974). Electrical Networks Containing Controlled Switches. IEEE
Symposium on Circuit Theory, pp.111
Kassakian, J.G.; Schlecht, M.F. & Verghese, G.C. (1991). Principles of Power Electronics,
AddisonWesley, ISBN:0201096897
Banerjee, S. & Verghese, G.C. (2001). Nonlinear Phenomena in Power Electronics, IEEE Press,
Piscataway, NJ, ISBN:0780353838
Orabi, M. & Ninomiya, T. (2003). Nonlinear Dynamics of PowerFactorCorrection Converter.
IEEE Transactions on Industrial Electronics, pp.11161125
Dranga, O.;Tse, C.K. & Siu Chung Wong (2005). Stability Analysis of Complete TwoStage
PowerFactorCorrection Power Supplies Proc. European Conference on Circuit Theory
and Design, pp.I/177I/180
Mohler, R.R. (1991). Nonlinear Systems, Vol.2, Applications to Bilinear Control. Prentice Hall,
Englewood Cliffs, NJ, ISBN:0136235212
Escobar, G.; Chevreau, D.; Ortega, R. & Mendes, E. (1999). An Adaptive PassivityBased
Controller for a Power Factor Precompensator. Proc. 5th European Control Conference,
BP41
Escobar, G.; Chevreau, D.; Ortega,R. & Mendes, E. (2001). An Adaptive PassivityBased
Controller for a Unity Power Factor Rectier. IEEE Trans. Control Systems Technology,
Vol.9, No.4, pp.637644
Sasaki, S. & Uchida, K. (1998). Nonlinear H
2
g
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 41
will settle into a periodic orbit inside the box provided the vector field is time independent.
When there exists time dependence, a bounded motion (which can be periodic or quasi
periodic) is produced inside the box. This motion is equally robust as the time independent
case, confining the centroid of the drop to any prescribed region provided the appropriate
incoming flow can be produced at the boundary.
In Fig. 2a, a rectangle is shown about the stagnation point of the flow field. The boundaries
of this rectangle serve as the limits where the position of the center of mass of the fluid
particle is allowed to stay at the nominal flow conditions. Fig. 2b shows a detailed sketch of
the tolerance domain PQRS above and next to the nominal stagnation point. The dark
streamlines correspond to the nominal flow field with outgoing flow exiting through the
side QO
R
and entering at the side RO
L
of the tolerance area. The dashed flow lines which
correspond to the shifted stagnation point y
SS
with the flow field essentially reversed: the
flow enters the PQ side and exists via RS. Also the flow lines which correspond to y
SS
behave in a symmetric manner relative to the symmetric tolerance area.
Fig. 2. Streamlines around the stagnation point of the unperturbed flow field generated by a
TRM.
When a fluid particle is placed around the stagnation point and the flow is started, the
particles centroid drifts away. Assume the centroid is initially at the position A at time t =
0 located inside the tolerance area shown in Fig. 3. In this position, the flow corresponds to
the nominal conditions, and the centroid is subsequently advected along the outgoing
direction reaching B at t = t
on
when the control is applied. The control scheme effect is to
displace the stagnation point to y
SS
, switching the flow field at B to one towards the nominal
stagnation point. As a result, the centroid follows the flow lines along the path BC, arriving
at C at time t = t
off
. At t
off
the flow is reset to the nominal conditions and the stagnation point
is moved back to the y
Nom
position; thus, the centroid follows the path CD. At D the situation
is repeated but now shifting the stagnation point to  y
SS
until the centroid reaches E where
the stagnation point is shifted back to the nominal value and the centroid moves towards F
where the process is repeated. It is to be noted that the transit time spent along paths AB,
CD, EF is much longer that the transit time along sections BC and DE because of the small
eigenvalues associated with the corresponding incoming directions which produce the
motion shown. The slope of l
in
can be modified in order to adjust the instant when the flow
is reset to the nominal conditions.
r
d
P
Q R
S
O
R
O
L
y
Nom
y
ss
(a) (b)
Applications of Nonlinear Control 42
Fig. 3. Trajectory ABCDEF followed by the centroid of a fluid particle in the controlled flow
field. The control area is shown in grey. The nominal flow corresponds to the darker
continuous lines, and the dashed lines show the relative displacement of the flow field
during the controlled portion of the cycle. The angle between the incoming and outgoing
streamlines at the nominal stagnation point
Nom
and the angle at the corrective flow
ss
have essentially the same values.
The purpose of the implemented control scheme is to produce always an incoming flow for
the drop at the boundary of the tolerance area. Thus the center of mass is effectively moved
as a dynamical system with an unstable rest point (the stagnation point) but with an
incoming vector field in an area surrounding the box. This arrangement guarantees the
existence of an incoming vector field that provides a fixed periodic solution by the Poincar
Bendixson theory; see Ross (1984). In this case since the incoming vector field is time
dependent, a bounded trajectory is obtained that is approximately periodic. This nonlinear
procedure of balancing the repulsion at the critical point with the correction of the boundary
of the tolerance region always produces a very robust bounded trajectory inside any
prescribed area.
All displacements of the stagnation point are assumed to be carried out on a time scale small
compared to the dynamics of the drop. In the theoretical description above, both the
centroid of the drop and the streamlines are adjusted instantaneously. For a laboratory
experiment this will not be the case: the exact position of the centroid is determined after
processing the flow field images, determining the centroid position and then the driving
motors modify accordingly the flow field within a finite response time. The latter time lags
are taken into account in order to calculate the effect of the control in experimental
situations.
The relevant times involved are associated to the velocity of the video system fps and
the time of capture and processing of all images, the finite response time of the cylinders
to readjust their velocity as a consequence of the control, mainly due to the inertia of the
y
ss
y
ss
y
Nom
A
B
C
D
E
F
l
in
l
in
Q
o
Q
o
Q
i
Q
i
P
Q R
S
O
R
O
L
Nom
ss
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 43
mechanical system, and the time of response of the fluid around the drop to the
adjustment in the velocity of the TRM. The time of adjustment of the flow can be
estimated as the diffusion time =l/ based on the gap g between the cylinders and the
kinematic viscosity . The total response time + + =
c
must be smaller than the
characteristic time
d
of the internal motion of the drop which is a function of the Capillary
number and the viscosity ratio. In the present work, + given the material
properties of the fluids and geometry of the setup; Stokes solution implies such time scales
as well.
To determine the adjusted velocity field i.e., reposition the stagnation point position y
SS
(or
y
SS
) needed to ensure an incoming flow along PQ  a new flow line (assumed straight) is
calculated, entering at PQ and ending at y
SS
; see Fig. 3. This requirement gives y
SS
as a
function of the size of the control area. Parametrizing the flow, relating the position of the
stagnation point with the angular velocities of the rollers, and , maintaining II
2D
constant, the required values of
c
and
c
needed to move the position of the stagnation
point to y
SS
can be calculated.
The actual control step is modeled as follows: When the centroid reaches the boundary of
the tolerance area at B in Fig. 3, the angular velocities are readjusted to the calculated values
c
and
c
according to the ramp function
() =
,
,
2
1 +
(1)
where t
c
is the time when the centroid of the particle reaches the point B, and i takes the
values 1 or 2. The following step at point D is calculated in the same manner. It is remarked
that during the control steps the flow parameter II
2D
is kept constant while the change in is
less than 0.5% for all cases.
This control is different from that of Bentley and Leal. In detail, this control does not aim to
stabilize the center of mass of the drop at the stagnation point as in the proportional control.
The present control takes advantage of the knowledge of the local flow field and balances
the unstable motion at the stagnation point with a time dependent incoming flow at the
boundary, giving an effective dynamical system with a periodic or quasiperiodic orbit for
the centroid. This is achieved by repositioning the velocity field; thus, producing a bounded
motion inside the tolerance area as a result of a balance of instability and the modified flow.
In contrast, the control by Bentley and Leal modifies the flow to project the trajectory of the
centroid along the stable direction at the nominal stagnation point; essentially, a
proportional method attempts to trace the centroid of the drop back to the stagnation point.
3. Experimental setup
The TwoRoll Mill experimental setup is shown in Fig. 4. The setup consists mainly of the
TwoRoll Mill flow cell with controlled temperature, the driving system, the optical system
and the interface system. The flow cell consist of the main body, a set of rollers of different
radii with machining tolerances for cylinder's diameters and gap of less than 5 m. The
parallelism and eccentricity of the rollers axes is limited to less than 5 m , for a top to bottom
distance of 10 cm. The driving system consists of two servomotors Kollmorgen AKM11B with
Applications of Nonlinear Control 44
Fig. 4. Experimental setup. The flow cell with the motors and the optical system are shown.
their controllers SERCOS Servostar 300. The optical systems is made up with a Navitar
microscope with a telecentric objective with a motorized magnification of 12X, an adapter
Navitar 161390 with a magnification of 2X, and an IEEE 1394 CCD camera, model XCD
X700, by Sony set for a capture rate of 15 and 30 fps. The visual field is about 1.2 mm
lengthwise covered with 1024 by 768 pixels. The optical resolution (based on the real part of
the Optical Transfer Function) of the full assembly is about 8 m or better than 64 line
pairs/mm. The main system is mounted on a pneumatically levitated optical workstation by
Newport Research.
The typical observed response time for the computer interface, power electronics and
cylinder's inertia is less than 0.01 seconds, for changes of rotational speeds less than 5% of
the preset values. The flow parameters and the position of the stagnation point are adjusted
by varying simultaneously the angular velocities of both cylinders and keeping II
2D
constant.
Figure 5 shows a schematic block diagram of the experimental setup. A drop is initially
placed near the stagnation point of the flow cell. The optical system captures images of the
drop and in the computer, the shape and the center of mass is calculated, and the angular
velocities of the rollers are calculated in order to maintain the drop in the position desired.
The calculated angular velocities are sent to the motors controllers and a new image of the
drop is captured and the cycle is repeated.
The interface system consists of a workstation HP XW4300 with a PCI SERCOS expansion
card, which communicates with the motor controllers. The control software is programmed
in Visual C++, in realtime mode. A Graphical User Interface (GUI) is used to provide access
to the functionality of the application, see Fig. 6. It incorporates two different aspects that
can be manipulated separately, one concerning to the video and the other concerning to the
control of the motors.
For the video aspect, the GUI window has two displays, one is for the video input, that
shows the frame that is acquired by the camera; and the other display is for the processed
image, showing the contour of the drop and its center of mass, along with the tolerance area
(fixed with the slide bar in the window) and the lines that corresponds to the ingoing and
outgoing streamlines of the nominal stagnation point.
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 45
Fig. 5. Experimental block diagram.
Fig. 6. Graphical User Interface (GUI) of the TwoRoll Mill Experiment.
For the control of the motors, the GUI has two parts, one is for the manual control and the
other is for the automatic control. The manual control window is used for the initial
positioning of the drop around the nominal stagnation point, inside the tolerance area.
For the control of the motors the window has slide bars that allow controlling the velocities of
each cylinder as well as its direction of rotation. In order to place the drops in the center of the
Applications of Nonlinear Control 46
image, it is necessary to put it in that location; this is done by manipulating its trajectory with
the controls above mentioned. Once the drop is in the right place (it means the center of mass
is inside the tolerance area), the automatic control is activated and the experiment goes on. The
monitoring section in the window allows us to watch the instantaneous velocities of the
motors, the coordinates of the center of mass and the size of the tolerance area.
4. Control scheme implementation
The automatic control for the experiment requires a realtime process. This program consists
basically of three parts: (i) The image acquisition (ii) Image analysis (iii) Calculus and
adjustment of the velocities of the motors.
4.1 Image acquisition and analysis
The images are provided by the CCD camera using the Instrumental and Industrial Digital
Camera Application Programming Interface (IIDCAPI) by Sony in the C++ program to
handle the frames provided by the camera. This section consists in two parts, the first one
take the frame and stored it in a file, the second one creates a list constantly updated in
order to always have the last image acquired by the camera available for the analysis section
and in case of some delay in acquiring, have a reservoir of frames.
To carry out the analysis of the last frame taken, in order to find the center of mass of the
drop (in pixels) we use the Open Source Computer Vision Library (OpenCV). The principal
parameter is the threshold value which can be adjusted ranging from 0 to 255 in a gray
scale, used to make a binary image. In the binary image, the contour of the drop is
computed using the Canny algorithm. Then the center of mass is found using the
corresponding discretized integral.
4.2 Calculus and adjustment of the velocities of the motors
Is in this part where the control takes place. Once the program has the coordinates of the
center of mass of the drop, decides whether is inside the tolerance area. If is not, the
program calculates and modifies the velocities of the motors depending on the position of
the center of mass of the drop. Also, in this part, the data of the number of the frame
processed, the coordinates of the center of mass and the size of the tolerance area are stored
in a file. The diagram of the control scheme is shown in Fig. 7.
Fig. 7. Diagram of the control scheme.
Image analysis (Position of
center of mass)
Selection of type
of flow (nominal)
Size of tolerance
area
Data
acquisition
Adjust of the
motor velocities
Movement of the
drop
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 47
5. Experimental results
5.1 Parameter used in studies of drop deformation
The parameters that govern the drop deformation and breakup are the ratio of the drop
viscosity to that of the suspending fluid
=
II
(2)
Where II
(3)
Thus is given by
=
(4)
Where
is the objective vorticity tensor which measures the rate of rotation of a material
point with respect to the rate of deformations principal axes at that point.
A dimensionless measure of the magnitude of the drop deformation is needed. This
parameter is the Taylor Deformation Parameter D
T
, defined in terms of the longest and
shortest semiaxes of the ellipsoidal drop cross section, see Fig 8.
Fig. 8. Deformation Parameter DT and orientation angle.
(5)
rmin
rmax
a
Applications of Nonlinear Control 48
The orientation angle of the drop is the angle between the longest axis of the drop and the x
axis.
The flowtype parameter value is related with the angles between the incoming and the
outgoing axes shown in Fig. 3 as
= 2
and as
= 2
.
In the experiment, the outgoing flow direction, about the stagnation point, is always
observed responding to small differences of the refractive index of Fluid 1 due to very small
differences of temperature between fluid volumes around each cylinder. Based upon
reversal of the flow field counter and corotating directions of the cylinders, the angle
between incoming and outgoing flow can be accurately measured within 0.1 degrees inside
the visual field less than 1.2 mm lengthwise. The observed angles are correct within 0.05
degrees of the nominal values here presented. This small angular uncertainty implies
uncertainties for the nominal flowtype parameter of less than 0.5%. Thus, the
experimental flow obtained is an excellent approximation for the theoretical one predicted.
The exterior fluid (Fluid 1) is a Polydimethylsiloxane oil DMS 25, = 485 mPa s, with a
relative density of 0.971. A drop (Fluid 2) of vegetable canola oil, filtered through a 3 m
pore size. At 23, the viscosity of Fluid 2 is = 72.6 mPa s, with a relative density of 0.917 is
used. Both liquids have a well defined Newtonian behavior at the interval of shear rate
values used. The following figures show the effect on the deformation, orientation and
trajectories of the centroid of a drop due to variations on the l
in
parameter. This parameter
permit adjusts the control in order to decrease the drifting effects of the
c
time, which is a
nonlinear function of the viscosity of the surrounding fluid. The drop tested had a diameter
of 1.0mm, Ca = 0.1031.
Fig. 9. Deformation, orientation and trajectory of the centroid of the drop, using the
parameter l
in
= 20. The mean deformation is D
T
= 0.1039, STD=0.002 and the mean
orientation angle is 41.8.
0
20
40
60
80
100
120
0
0.02
0.04
0.06
0.08
0.1
0.12
0 10 20 30 40 50 60 70
O
r
i
e
n
t
a
t
i
o
n
A
n
g
l
e
[
D
e
g
]
D
e
f
o
r
m
a
t
i
o
n
P
a
r
a
m
e
t
e
r
t [s]
Deformation and Orientation of the drop
0.1
0.06
0.02
0.02
0.06
0.1
0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 49
Fig. 10. Deformation, orientation and trajectory of the centroid of the drop, using the
parameter l
in
= 30. The mean deformation is D
T
= 0.1042, STD=0.0021, the mean orientation
angle is 41.8.
Fig. 11. Deformation, orientation and trajectory of the centroid of the drop, using the
parameter l
in
= 40. The mean deformation is D
T
= 0.1045, STD=0.0025, and the mean
orientation angle is 41.8.
0
20
40
60
80
0
0.02
0.04
0.06
0.08
0.1
0.12
0 10 20 30 40 50 60
O
r
i
e
n
t
a
t
i
o
n
A
n
g
l
e
[
D
e
g
]
D
e
f
o
r
m
a
t
i
o
n
P
a
r
a
m
e
t
e
r
t [s]
Deformation and Orientation of the drop
0.08
0.03
0.02
0.07
0.35 0.25 0.15 0.05 0.05 0.15 0.25 0.35
0
20
40
60
80
100
0
0.02
0.04
0.06
0.08
0.1
0.12
0 10 20 30 40 50 60
O
r
i
e
n
t
a
t
i
o
n
A
n
g
l
e
[
D
e
g
]
D
e
f
o
r
m
a
t
i
o
n
P
a
r
a
m
e
t
e
r
t [s]
Deformation and Orientation of the drop
0.08
0.03
0.02
0.07
0.35 0.25 0.15 0.05 0.05 0.15 0.25 0.35
Applications of Nonlinear Control 50
Fig. 12. Deformation, orientation and trajectory of the centroid of the drop, using the
parameter l
in
= 50 . The mean deformation is D
T
= 0.1049, STD=0.0027, the mean orientation
angle is 41.8.
Fig. 13. Deformation, orientation and trajectory of the centroid of the drop, using vertical
limits. The mean deformation is D
T
= 0.1039, STD=0.0020, the mean orientation angle is 41.8.
0
20
40
60
80
100
0
0.02
0.04
0.06
0.08
0.1
0.12
0 10 20 30 40 50 60 70
O
r
i
e
n
t
a
t
i
o
n
A
n
g
l
e
[
D
e
g
]
D
e
f
o
r
m
a
t
i
o
n
P
a
r
a
m
e
t
e
r
t [s]
Deformation and Orientation of the drop
0.08
0.03
0.02
0.07
0.35 0.25 0.15 0.05 0.05 0.15 0.25 0.35
0
20
40
60
80
100
0
0.02
0.04
0.06
0.08
0.1
0.12
0 10 20 30 40 50 60
O
r
i
e
n
t
a
t
i
o
n
A
n
g
l
e
[
D
e
g
]
D
e
f
o
r
m
a
t
i
o
n
P
a
r
a
m
e
t
e
r
t [s]
Deformation and Orientation of the drop
0.12
0.08
0.04
0
0.04
0.08
0.12
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 51
6. Numerical results and comparisons with the experimental results
In order to study numerically the evolution of the drop on the proposed control strategy, we
use the results of Reyes et al. (2011). We just outline the main steps to introduce the
extensions used in this Chapter to study the effect of noise and the imperfections present in
the experimental situation.
We thus proceed as in Reyes et al. (2011) advancing the drop boundary using the equation
= (
)
(6)
where the velocity
.
t
0.3
0.2
0.1
0.0
0.1
0.2
0.3
X coordinate
5 10 15 20 25 30
.
t
0.04
0.02
0.00
0.02
0.04
0.06
Y coordinate
Nonlinear Control Applied to the Rheology of Drops in Elongational Flows with Vorticity 55
Beyond this time, the experimental data shows larger excursions. These are due to the
mismatch between the commercial worm gear and worm mechanism used to reduce the
angular velocity of the motors and transmit the motion to the rollers.
It is to be noted the remarkable agreement in the actual drop parameters which are the ones
of interest.
7. Conclusions
We have shown that it is possible to maintain the position of a drop about the unstable
stagnation point of the flow field generated by a TRM setup. This scheme is of the upmost
importance for studies of drops in elongational flow with significant vorticity, and of
relevance because its space of parameters is not accessible to FRM flows previously studied
since Bentley (1986a). Indeed, TRMs expand the family of 2D elongational flows amenable
with a four roll mill, albeit the former flows carry amounts of vorticity similar to that of
simple shear flows.
But because the TRMs configuration can only displace the stagnation point along the line
between the cylinder axes, the control scheme developed for FRMs or PBAs cannot be used
for studies of drop's dynamics in elongational flows of the TRM type. For TRMs the control
scheme is based upon features of PoincarBendixson limiting cycles. However, these limit
cycles do not imply that the wandering trajectory of the drop has to be contained in a very
tight region about the nominal stagnation point. The control scheme appears capable of
working for a tolerance region with the cases tested experimentally.
The control implemented in the experimental has shown to be successful. At this point,
several complications have been seen in the implementation: nevertheless the control
scheme is robust enough to keep the drop inside a region where the parameters of interest
have a low variation, for a long times, enough to have reliable measures of the relevant
parameters. The figures 913 shown that despite the trajectory of the drop, the parameters of
interest in drop dynamics (D
T
and orientation angle) are the same.
It is important to mention that even when the comparisons are made just for only one flow
type parameter, the results shown that it is reasonable to expect the same when we will use a
different flowtype parameter, i.e. a different geometry.
Important differences have been observed between the experimental and numerical
trajectories, this disagreement is due to mechanical imperfections, and could be fixed in the
future by using an system motion transmission with better precision.
8. Acknowledgment
IYR thanks to CEPUNAM for funding his graduate fellowship, MAHR thanks DGAPA
UNAM, AAM thanks the FENOMEC program. EG thanks CONACyT research grants.
9. References
Acrivos, A. (1983). The breakup of small drops and bubbles in shear flows. Ann. NY Acad.
Sci., 404:111
Applications of Nonlinear Control 56
Arfken, G. (1971). Mathematical Methods for Physicists. 2nd Ed. Academic Press
Astarita, G. (1979). Objective and generally applicable criteria for flow classification. Journal
of NonNewtonian Fluids Mechanics, 6:6976
Bentley, B. J. and L. G. Leal. (1986a). A computercontrolled fourroll mill for investigations
of particle and drop dynamics in twodimensional linear shear flows. J. Fluid Mech.
167:219240
Bentley, B. J. and L. G. Leal. (1986b). An experimental investigation of drop deformation and
breakup in steady twodimensional linear flows. J. Fluid Mech. 167:241283
Birkhoffer, Beat H. et al. (2005). Computercontrolled flow cell for the study of particle and
drop dynamics in shear flow fields. Ind. Eng. Chem. Res. 44:69997009
Geffroy E. and L. G. Leal. (1992). Flow Birefringence of a Concentrated Polystyrene Solution
in a Two RollMill 1. Steady Flow and StartUp of Steady Flow. J. Polym. Sci. B:
Polym. Phys. 30(12):13291349
Gradshteyn, I. S. and I. M. Ryzhik. (1981). Tables of integrals, series and products. Academic
Press
Jeffery, G. B. (1922). The rotation of two circular cylinders in a viscous fluid. Proc Royal
Society A 101:169174
Pozrikidis, C. (1992). Boundary Integral and singularity methods for linearized viscous flow.
Cambridge University Press
Pozrikidis, C. (2003). Modeling and simulation of capsules and biological cells. CRC Press
Rallison J. M. (1980). A note of the timedependent deformation of a viscous drop which is
almost spherical. J. Fluid Mechanics 98(3):625633
Reyes, M. A. H. and E. Geffroy. (2000). Study of low Reynolds number hydrodynamics
generated by symmetric corotating tworoll mills. Revista Mexicana de Fsica
46(2):135147
Reyes, M. A. H. and E. Geffroy. (2000). A corotating tworoll mill for studies of two
dimensional, elongational flows with vorticity. Phys. Fluids 12(10):23722376.
Reyes, M. A. H., A. A. Minzoni & E. Geffroy (2011), Numerical study of the effect of
nonlinear control on the behavior of a liquid drop in elongational flow with
vorticity, Journal of Engineering Mathematics, 71(2):185203.
Ross, S. (1984). Differential Equations. Wiley
Singh, P. and L. G. Leal. (1994). Computational studies of the FENE dumbbell model in a co
rotating tworoll mill. J. Rheol. 38:485517
Stone, H. A. (1994). Dynamics of drop deformation and breakup in viscous fluids. Ann. Rev.
Fluid Mech. 26:65102
Taylor, G. I. (1932). The Viscosity of a Fluid Containing Small Drops of Another Fluid. Proc.
R. Soc. London, Ser. A, 138:4148
Taylor, G. I. (1934). The Formation of Emulsions in Definable Fields of Flow. Proc. R. Soc.
London, Ser. A 146:501523
Torza S., R. G. Cox and S. G. Mason. (1972). Particle motions in shear suspensions. J. Colloid
and Interface Sci. 38(2):395411
Wang, J. J., D. Yavich, and L. G. Leal. (1994). Time resolved velocity gradient and optical
anisotropy in lineal flow by photon correlation spectroscopy. Phys. Fluids
6(11):35193534
Yang, H., C. C. Park, Y. T. Hu, and L. G. Leal. (2001). The coalescence of two equalsized
drops in a two dimensional linear flow. Phys. Fluids 13(5):10871106
4
1
Robust Control Research
of Chaos Phenomenon for
DieselGenerator Set
on Parallel Connection
Manlei Huang
School of Automation, Harbin Engineering University, Harbin,
China
1. Introduction
Several dieselgenerator sets usually operate on parallel connection in ship power system,
which has altitudinal nonlinearity. When operating point of system changes, its dynamic
property will change markedly. The oscillation phenomenon of ship power system which is
acyclic, random and gusty or paroxysmal will occur on light load working condition, it can
result in system sectionalizing when it is serious, this phenomenon is called chaos. Chaos is
a very complicated phenomenon which is generated by the interaction of each parameter in
the nonlinear system. When it appears in ship power system, following the continuous and
random oscillation of system operating parameter, which endangers operation security of
system seriously, it must be prevented and eliminated effectively in the system. In order to
analyze the chaos phenomenon of ship power system, the nonlinear mathematical model of
two dieselgenerator sets on parallel connection is built in this paper, which reflects the
variation law of ship power system. Then the light load working condition of two diesel
generator sets on parallel connection in ship power station is analyzed by using Lyapunov
index method on the base of this, seeking the generating mechanism of chaos. A nonlinear
robust synthetic controller is designed which is based on the nonlinear mathematical model
of dieselgenerator set, then a nonlinear robust synthetic control law is developed for the
dieselgenerator set, it will be applied to control the chaos phenomena, thus providing
desirable stability for ship power system.
2. Mathematic model of dieselgenerator set on parallel connection
The mathematical model of dieselgenerator set include the mathematical model of
electromechanical transient process and electromagnetism transient process, first building
the mathematical model of electromechanical transient process, then the mathematical
1
Foundation item: Supported by the National Natural Science Foundation of China under Grant
No.60774072; Fundamental Research Funds for the Central Universities of China under Grant
No.HEUCFT1005.
Applications of Nonlinear Control
58
model of electromagnetism transient process, the mathematical model of one diesel
generator set is get on the base of this.
The mathematical model of dieselgenerator set electromechanical transient process
describes the motion law of dieselgenerator set
[13]
, reflecting the dynamic change process of
power angle and angular speed, its expression is
2
2
1
1
1 1 1
2
2
d
d
d
d
0 0 0 0 0
,
sin sin .
q d q
b
a a a a d a d q
t
E U X X
T c U
c L
t T T T T X T X X
o
e
e
e o o
e e e e e
' '
= + +
' '
(1)
In the equation: o is power angle of dieselgenerator set, e electric angular speed of diesel
generator set, U terminal voltage of generator stator winding,
q
E' qaxis transient electric
potential, X reactance of each winding, L output axis displacement of diesel engine
govonor actuator, rad/s
0
100 e = ,
a
T ,
b
T ,
1
c ,
2
c constants, o , L real values, other variables
per unit values.
From Eq.(1) we know, this equation has nonlinear feature.
The mathematical model of dieselgenerator set electromagnetism transient process include
stator voltage balance equation of generator and electromagnetism transient equation of
field winding
[3]
, omitting the effect of damping winding, its expression is
0 0 0
2 2
1 1
d
d
,
,
,
.
q
d d
fd q d
d d d
d d q q
q q d d q
d q
E
X X
E E I
t T T T
U RI X I
U RI X I E
U U U
e
e e
'
'
' =
= +
' ' = +
= +
(2)
In the equation: U is terminal voltage of stator winding,
d
U and
q
U daxis and qaxis
component of stator winding terminal voltage, R resistance of stator winding, X reactance
of each winding, I current of each winding, T time constant of each winding,
q
E' qaxis
transient electric potential,
fd
E voltage of exciting winding.
Since
q
E' is not easy to measure, we select the terminal voltage of stator winding U as state
varible, only needing change
q
E' of Eq.(2) into U . According to the relation between varible
data, the following form is set up
3 q
U E c o ' =
(3)
Substituting Eq.(3) into the first item Eq.(2), we have
0 0 0
1 1 d
d
3
3 3
0
d d
fd d
d d d d
c X X U
E U c I c
t T T T T
o e
'
= +
(4)
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
59
Substituting Eq.(3) into Eq.(1) and combining with Eq.(4), we have
2
2
1
0 0 0
1
1 1 1
2
2
1 1
d
d
d
d
d
d
3
0 0 0 0 0
3
3 3
0
,
( )
sin sin ,
d q
b
a a a a d a d q
d d
fd d
d d d d
t
X X
T U c U c U
c L
t T T T T X T X X
c X X U
c U E I c
t T T T T
o
e
o e
e o o
e e e e e
o e
'
+
= + +
' '
'
= + +
(5)
We know from the expression of current
d
I
cos
q
d
d
E U
I
X
o '
=
'
(6)
Substituting Eq.(3) into Eq.(6), we have
3
cos
d
d
U c U
I
X
o o +
=
'
(7)
Substituting Eq.(7) into Eq.(5), we get
2
2
1
0 0
1
1 1 1
2
2
1
d
d
d
d
d
d
3
0 0 0 0 0
3
3 3
0 0
,
( )
sin sin ,
cos
d q
b
a a a a d a d q
d d d d
fd
d d d d d d d
t
X X
T U c U c U
c L
t T T T T X T X X
X c X X X U
c U E U c
t T X T X T T X
o
e
o e
e o o
e e e e e
o e o
'
+
= + +
' '
'
= + + +
' ' '
(8)
Eq.(8) is the nonlinear mathematical model of one dieselgenerator set, which reflects the
relationship of interaction and mutual influence among power angle, speed and voltage,
describing the variation law of three variables more exactly.
We select dq axis of first synchronous generator as reference frame, building the
mathematical model of two dieselgenerator sets on parallel connection. Suppose two diesel
generator sets have the same power, type and parameters, d,q component of load current is
d
I ,
q
I , power anger difference of two synchronous generators is
12
o , the mathematic model
of dieselgenerator set on parallel connection is
2
2
1
0 0
1
1 1 1
2
2
1
d
d
d
d
d
d
3
0 0 0 0 0
3
3 3
0 0
,
( )
sin sin ,
cos
i
i
d q
i b i i i i
i i i i
a a a a d a d q
i d d d d
i i i fdi i i
d d d d d d d
t
X X
T U c U U c
c L
t T T T T X T X X
U X c X X X
c U E U c
t T X T X T T X
o
e
e o
e o o
e e e e e
o e o
'
+
= + +
' '
'
= + + +
' ' '
(9)
Applications of Nonlinear Control
60
In the equation: 1, 2 i = , subscript 1 indicates the first dieselgenerator set, subscript 2
indicates the second dieselgenerator set.
Current coupling relation of two dieselgenerator sets is
1 2 12 12
1 2 12 12
cos sin
sin cos
d d d
q q q
I I I
I I I
o o
o o
( ( ( (
= +
( ( ( (
(10)
Eq.(10) describes the current distribution relation after two dieselgenerator sets enter into
parallel connection.
Voltage coupling relation of two dieselgenerator sets is
1 2 12 12
1 2 12 12
cos sin
sin cos
d d
q q
U U
U U
o o
o o
( ( (
=
( ( (
(11)
Eq.(11) describes the voltage restriction relation after two dieselgenerator sets enter into
parallel connection.
Eq.(9), Eq.(10) and Eq.(11) are the nonlinear mathematical model of two dieselgenerator sets
on parallel connection, which reflects the relationship of interaction and mutual influence
between two dieselgenerator sets, describing the variation law of power angle, speed and
voltage on two dieselgenerator sets exactly.
3. Chaos oscillation analysis of dieselgenerator set on parallel connection
This paper will make research on the stability of ship power system by the variation law of
power angle, speed and voltage on dieselgenerator set. Due to the power transmission
between the dieselgenerator sets on parallel connection, it is apt to produce power
oscillation. Power oscillation is the dynamic process of power of dieselgenerator set
regulating repeatedly under the effect of some periodic interference.
Power oscillation is a chaos oscillation in nature from the point of view of chaos. The
fundamental feature of chaotic motion is highly sensitive to initial condition, the track which
is produced by two initial values which is very near each other, will separate according to
index pattern as time elapses, Lyapunov index is the quantity which describes the
phenomenon. Distinguishing methods of time series chaotic character include fix quantity
analysis and ocular analysis, first making numerical analysis for Lyapunov index and
judging the condition of chaos emerging, then determining if the chaos exists or not under
this condition by the method of ocular analysis. The methods of ocular analysis include time
course method, phase path chart method, strobe sampling method, Poincare cross section
method and power spectrum method. The methods of calculating Lyapunov index include
definition method, Wolf method and Jacobian method, Jacobian method is a method of
calculating Lyapunov index which develops in real application. This paper will use Jacobian
method to calculate Lyapunov index.
Considering following differential equation system
( ) x F x = (12)
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
61
In the equation:
dx
x
dt
= ,
m
x R e . The evolution of tangent vector e of dot ( ) x t in the tangent
space can be expressed by the equation as follow
( ( )) e T x t e = ,
F
T
x
c
=
c
(13)
In the equation: T is Jacobian matrix of F . The solution of equation can be expressed as
( ) ( , (0)) e t U t e = (14)
In the equationU : (0) ( ) e e t is mapping of linear operator. The asymptotic behavior of
mapping U can be described by index as
( ) 1
( (0), (0)) lim ln
(0)
t
e t
x e
t e
=
(15)
So, the Lyapunov index of system (12) can be formulated as the mean of above repeat
process
1
(( 1) )
1
lim ln
( )
(( 1) ) ( ) (2 ) 1
lim ln
( ) (( 1) ) ( )
k
k
j
k
e j t
k t e j t
e k t e k t e t
k t e k t e k t e t
+ A
=
A A
+ A A A
=
A A A A
(16)
For a phase space of n dimension, there will be n Lyapunov index, arranging them
according to the order from big to small, supposing ( )
1 2 n
> > > ,
1
is called
maximum Lyapunov index. Generally speaking, having negative Lyapunov index
corresponds with contracting direction, the tracks which are near are stable in the part,
corresponding periodic motion. The positive Lyapunov index indicates that the tracks
which are near separate by index, the strange attractor is formed in phase space, the
Lyapunov index is bigger, the chaotic nature of system is stronger, vice versa. For a phase
space of n dimension, the maximum Lyapunov index is whether bigger than 0 or not is the
basis of judging the system if has chaos oscillation or not.
Computing Lyapunov index of two dieselgenerator sets on parallel running with light load
separately, two dieselgenerator sets all use conventional controllers. Fig.1 and Fig.2 give the
phase diagram of power angle, speed and voltage of two dieselgenerator sets on parallel
running with 12.5% load separately.
Two dieselgenerator sets run for 100 seconds on parallel connection with 12.5% load, the
initial va1ue of No.1 dieselgenerator set is: ( , , ) (0.1017, 1.0662, 0.9762) U o e = , Lyapunov
index is:
1
0.076789 = ,
2
0.035235 = ,
3
0.197558 = ; the initial va1ue of No.2 diesel
generator set is: ( , , ) (0.1022, 1.0662, 0.9762) U o e = , Lyapunov index is:
1
0.076806 = ,
2
0.035230 = ,
3
0.197571 = .
Fig.3 and Fig.4 give the phase diagram of power angle, speed and voltage of two diesel
generator sets on parallel running with 25% load separately.
Applications of Nonlinear Control
62
40
30
20
10
0
10 0.5
0
0.5
1
1.5
0.8
1
1.2
1.4
1.6
1
1
U
1
Fig. 1. Phase diagram of No.1 dieselgenerator set when two sets load 12.5% on parallel
connection
40
30
20
10
0
10 0.5
0
0.5
1
1.5
0.8
1
1.2
1.4
1.6
2
2
U
2
Fig. 2. Phase diagram of No.2 dieselgenerator set when two sets load 12.5% on parallel
connection
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
63
15
10
5
0
5
0.7
0.8
0.9
1
1.1
0.9
1
1.1
1.2
1.3
1
1
U
1
Fig. 3. Phase diagram of No.1 dieselgenerator set when two sets load 25% on parallel
connection
15
10
5
0
5
0.7
0.8
0.9
1
1.1
0.9
1
1.1
1.2
1.3
2
2
U
2
Fig. 4. Phase diagram of No.2 dieselgenerator set when two sets load 25% on parallel
connection
Two dieselgenerator sets run for 100 seconds on parallel connection with 25% load, the
initial va1ue of No.1 dieselgenerator set is: ( , , ) (0.1812, 1.0607, 0.9686) U o e = , Lyapunov
index is:
1
0.079251 = ,
2
0.034251 = ,
3
0.199955 = ;the initial va1ue of No.2 diesel
Applications of Nonlinear Control
64
generator set is: ( , , ) (0.1820, 1.0607, 0.9686) U o e = , Lyapunov index is:
1
0.078311 = ,
2
0.034953 = ,
3
0.199742 = .
From Fig.1 to Fig.4 we can see, all maximum Lyapunov indexes of the system are greater than
0, showing that the system exists chaotic phenomenon. Two dieselgenerator sets with light
load on parallel connection, enter into chaotic state after running a length of time, their specific
expression are the oscillation of power angle and speed. Two dieselgenerator sets on parallel
connection load the lighter, the oscillation of power angle and speed is severer. The oscillation
of power angle means the oscillation of power, the reason is the nonlinearity of ship power
system and the power transmission between the two dieselgenerator sets. The controller in
this paper is proportional controller, which is a linear controller. It cant control the nonlinear
character of ship power system, it cant average the load in parallel operation control, there is a
power angular difference between the dieselgenerator sets, thus engendering the power
transmission between the sets, which results in the happening of chaotic phenomenon.
Fig.5 and Fig.6 give the phase diagram of power angle, speed and voltage of two diesel
generator sets on parallel running with 25% load plus periodicity load separately. The
periodicity load usually appears in ship power system, it is widespread.
Two dieselgenerator sets run for 100 seconds on parallel connection with 25% load increasing
periodicity load 0.01sint, the initial va1ue of No.1 dieselgenerator set is:
( , , ) (0.1812, 1.0607, 0.9686) U o e = , Lyapunov index is:
1
0.073257 = ,
2
0.031824 = ,
3
0.191497 = ; the initial va1ue of No.2 dieselgenerator set is:
( , , ) (0.1820, 1.0607, 0.9686) U o e = , Lyapunov index is:
1
0.073393 = ,
2
0.030161 = ,
3
0.189995 = .
30
20
10
0
10 0.2
0.4
0.6
0.8
1
0.8
1
1.2
1.4
1.6
1
1
U
1
Fig. 5. Phase diagram of No.1 dieselgenerator set when two sets increase periodicity load
meanwhile load 25% on parallel connection
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
65
30
20
10
0
10 0.2
0.4
0.6
0.8
1
0.8
1
1.2
1.4
1.6
2
2
U
2
Fig. 6. Phase diagram of No.2 dieselgenerator set when two sets increase periodicity load
meanwhile load 25% on parallel connection
From Fig.5 and Fig.6 we can see, all maximum Lyapunov indexes of the system are greater
than 0, showing that the system exists chaotic phenomenon. The oscillation of power angle
and speed in two dieselgenerator sets is severer than the state of not increasing periodicity
load. Because the periodicity load is nonlinear, increasing periodicity load intensifies the
nonlinearity of system, thus aggravating the power oscillation of system
[46]
.
The computer simulation results show that it exists chaotic oscillation phenomenon when
two dieselgenerator sets run on parallel connection with light load. Primary cause of
emerging this phenomenon is the nonlinearity of ship power system, minor cause is the
power transmission between the two dieselgenerator sets. Besides this, using conventional
linear controller is also a key factor of generating the chaotic oscillation of system. Only
using nonlinear controller, making the nonlinear characteristic of ship power system offset
and compensate, can we solve the problem of system chaotic oscillation fundamentally. The
chaotic oscillation phenomenon is transition state between stable state and unstable state, it
must be prevented in order to ensure the stability of system.
4. Design of nonlinear robust synthetic controller
Mixed Htwo/Hinfinity control theory is a robust control theory that has speed
development from the eighties of 20 century, which can solve the problem of robust stability
and robust performance
[710]
. Because dieselgenerator set control system is a nonlinear
control system, using the method of direct feedback linearization to linearize the nonlinear
system, the state feedback controller is designed for linearization system using mixed H
two/Hinfinity control theory, thus acquiring nonlinear robust control law in order to reach
the purpose of restraining the chaotic phenomenon of ship power system, improving the
stability of ship power system.
Applications of Nonlinear Control
66
Because of coupling action between speed and voltage, the nonlinear robust synthetic
controller is designed for dieselgenerator set in order to control speed and voltage
synthetically, making the both interaction in minimum range, thus improving the stability of
frequency and voltage in ship power system.
The principle diagram of dieselgenerator set synthetic control system based on nonlinear
robust synthetic controller is shown in Fig.7. The dieselgenerator set synthetic control
system is made up of diesel engine, generator, nonlinear robust synthetic controller,
actuator, oil feeding mechanism and exciter. Nonlinear robust synthetic controller is made
up of two parts, one is nonlinear Htwo/Hinfinity speed controller, another is nonlinear H
two/Hinfinity voltage controller.
The differential equation of actuator is
d
dt
1
1
1 1
L L K
u
T T
= + (17)
The differential equation of exciter is
d
dt
2
2
2 2
fd fd
E E
K
u
T T
= + (18)
Fig. 7. Principle diagram of dieselgenerator set synthetic control system
First step, we design nonlinear Htwo/Hinfinity speed controller.
Combining Eq.(1) with Eq.(17), we can get the nonliear mathemaical model of diesel engine
speed regulation system.
2
2
1
1
1 1 1
2
2
d
d
d
d
d
dt
0 0 0 0 0
1
1
1 1
,
sin sin ,
q d q
b
a a a a d a d q
t
E U X X
T c U
c L
t T T T T X T X X
L L K
u
T T
o
e
e
e o o
e e e e e
' '
= + +
' '
= +
(19)
Since Eq.(19) has nonlinear feature, the method of direct feedback linearization is used to
linearize Eq.(19). Order
1
x o = ,
2
1 x e = ,
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
67
2
2
1
1 1 1
2
2
3
0 0 0 0 0
sin sin
q d q
b
a a a a d a d q
E U X X
T c U
x c L
T T T T X T X X
e o o
e e e e e
' '
= + +
' '
, so Eq.(19) can be
written as
2
2 2
os2
1 2
2 3 1
1
3 3 1
0 1 0 1 0 0 0
( )
cos ( 1) ( 1)
q d q
b
a a a a d a d q
x x
x x d w
E U U X X
T c K c
x x u L c
T T T T T T X T X X
o e o e
e e e e e
= +
' '
= +
' '
(20)
In the equation:
1
d w is the undesired signal which is assumed for using Htwo/Hinfinity
control method, including equivalence disturbance which is generated by disturbance
torque and modeling error.
Assigning virtual controlled variable
2
2 2
os2
1
1
1 0 1 0 0 0
( )
cos ( 1) ( 1)
q d q
a a a d a d q
E U U X X
c K c
v u L c
T T T T T X T X X
o e o e
e e e e
' '
=
' '
(21)
So Eq.(20) can be changed as
1 2
2 3 1
3 3
0
b
a
x x
x x d w
T
x x v
T e
= +
= +
(22)
Eq.(22) can be written as matrix form
1 2
x Ax B w B v = + +
(23)
In the equation:
1
2
3
x
x x
x
(
(
=
(
(
,
0
0 1 0
0 0 1
0 0
b
a
A
T
T e
(
(
(
( =
(
(
(
,
1 1
0
0
B d
(
(
=
(
(
,
2
0
0
1
B
(
(
=
(
(
.
Defining the evaluation signal of dynamic performance as
1 11 12
2 2 21 22
z C x D w D v
z C x D w D v
= + +
= + +
(24)
In the equation:
11
1 12
13
0 0
0 0
0 0
q
C q
q
(
(
=
(
(
,
21
2 22
23
0 0
0 0
0 0
q
C q
q
(
(
=
(
(
,
11 21
0
0
0
D D
(
(
= =
(
(
,
12
1
0
0 D
r
(
(
=
(
(
,
Applications of Nonlinear Control
68
22
2
0
0 D
r
(
(
=
(
(
,
1
C ,
2
C ,
11
D ,
12
D ,
21
D ,
22
D are weighting matrix,
ij
q >0(i=1,2; j=1,2,3)and
i
r >0
(i=1,2) weighting coefficient. We can select optimal performance combination through
changing weighting coefficient, including stability of ship power system, frequency
regulation precision and low energy loss of speed regulation system.
For the control system made up of Eq.(23) and Eq.(24), requring design a controller F,
making the closed loop system asymptotically stable, moreover Hinfinity norm of closed
loop transfer function T
(s) from w to z
= +
'
= +
' '
= + +
' '
(29)
We select the error of voltage U A as state varible, The relation between U and U A is
0
U U U A = (30)
In the equation:
0
U is initial value of stator winding terminal voltage, its value is 1.
We use
e
P repalce
2
sin sin2
2
q d q
d d q
E U X X
U
X X X
o o
' '
+
' '
, regarding
e
P as external disturbance,
substituting Eq.(30) into Eq.(29), we get
0 0 0 0
2
1
1 1 1
1
1 1
d
dt
d
d
d
d
d
d
2
2
2 2
3
3 3
0
0 0 0 0
( )
fd fd
d d
fd d
d d d d d
b
e
a a a a
E E
K
u
T T
c X X U
E U c c I
t T T T T T
t
T c
c L P
t T T T T
o e
o
e
e
e
e e e e
= +
' A
= A +
= + +
(31)
Eq.(31) can be written as matrix form
1 2 2
x A x B w B u ' ' ' ' ' ' = + + (32)
In the equation:
fd
E
U
x
o
e
(
(
A
(
' =
(
(
(
,
2
3
3
0 0 0
0
1
0 0 0
1 1
0 0 0 1
0 0 0
d d d
b
a
T
c
c
T T T A
T
T e
(
(
(
(
(
' =
(
(
(
(
(
,
Applications of Nonlinear Control
70
3
0 0
1
2 1
0 0 0
0 0 0 0
1
0 0
0 0 0 1
1
0
d d
d d
a a a
X X
c
T T
B
c c
T T T e e e
(
(
'
(
(
' =
(
(
(
(
(
,
2
2
2
0
0
0
K
T
B
(
(
(
' ( =
(
(
(
,
1
d
e
I
L
w
P
(
(
(
' =
(
(
.
Defining the evaluation signal of dynamic performance as
1 11 12 2
2 2 21 22 2
z C x D w D u
z C x D w D u
(33)
In the equation:
14
15
1
16
17
0 0 0
0 0 0
0 0 0
0 0 0
q
q
C
q
q
(
(
(
' =
(
(
,
24
25
2
26
27
0 0 0
0 0 0
0 0 0
0 0 0
q
q
C
q
q
(
(
(
' =
(
(
,
3
4
12
5
6
r
r
D
r
r
(
(
(
' =
(
(
,
7
8
22
9
10
r
r
D
r
r
(
(
(
' =
(
(
,
11 21
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
D D
(
(
(
' ' = =
(
(
,
1
C' ,
2
C' ,
11
D' ,
12
D' ,
21
D' ,
22
D' are weighting matrix,
ij
q >0(i=1,2; j=4
7)and
i
r >0(i=310)weighting coefficient. We can select optimal performance combination
through changing weighting coefficient, including stability of ship power system, voltage
regulation precision and low energy loss of excitation system.
From Eq.(32) and Eq.(33), we can get the augmentation controlled object based on mixed H
two/Hinfinity control theory
1 2
1 11 12
2 21 22
A B B
P C D D
C D D
' ' '
(
(
' ' ' ' =
(
( ' ' '
(34)
For controlled object P' , existing Htwo/Hinfinity state feedback controller:
2 1 2 3 4 1 2 3 4
[ ]
fd
fd
E
U
u F x f f f f f E f U f f o e
o
e
(
(
A
(
' ' ' ' ' ' ' ' ' ' = = = + A + +
(
(
(
(35)
That is nonlinear Htwo/Hinfinity voltage control law of dieselgenerator set.
Third step, we design nonlinear robust synthetic controller.
Combining Eq.(28) with Eq.(35), we can get nonlinear robust synthetic controller of diesel
generator set
[1118]
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
71
1
1
2
2
2 os2
2
1 0 1 0 1 1 1
1 1 2 3 3 3 3
1 2 1 2 1 2 1 2 1 1 2 1
1 1
1
3
2 1 2 1 2 1
2 1 2 3 4
1
( 1) sin
( )
sin cos ( 1) ( 1)
q
a a b
d
d q q d q
d q d d q
fd
E U
T T T T T T T T T
u L f f f f c f L f
K c K c K c K c K K c K X
X X E UT U T X X
T U
f c
c K X X c K X c K X X
u f E f U f f
e e
o e e o
o o e o e
o e
'
= + + + + +
'
' ' '
+ +
(36)
Eq.(36) considers the coupling function of speed and voltage, which controls both
synthetically. It can increase dynamic precision of speed and voltage, improving the stability
of ship power system.
5. Results of computer simulation
The key parameters of dieselgenerator set control system in this paper are as follow:
The power of dieselgenerator set is 1250kW; the rated speed n =1500r/min; the rotary
inertia of set J=71.822kgm
2
; the damping coefficient of set D =5.54;the magnetic pole pair
number of generator p=2; the rated torque of diesel engine 11.9kNm; the maximum troke of
output axis 10mm.
The rated voltage of synchronous generator is 390V; the rated current 2310A; the power factor
0.8; the rated frequency 50Hz; the exciting voltage of exciter 83V; the exciting current 7.7A.
Designing nonlinear synthetic controller based on mixed Htwo/Hinfinity control theory,
assuming disturbance signal coefficient of Eq.(23)
1
d =0.1, assuming weighting coefficient of
Eq.(24)
11
q =0.002,
12
q =0.4,
13
q =0.5,
1
r =0.01,
21
q =0.002,
22
q =0.4,
23
q =0.5,
2
r =0.01.
Corresponding matrix are
0 1 0
0 0 1
0 0 0.0014
A
(
(
=
(
(
,
1
0
0.1
0
B
(
(
=
(
(
,
1 2
0.002 0 0
0 0.4 0
0 0 0.5
C C
(
(
= =
(
(
,
12 22
0
0
0.01
D D
(
(
= =
(
(
.
Using LMI toolbox, we get state feedback controller:
[ 0.3582 35.1042 157.7578] F = (37)
Assuming weighting coefficient of Eq.(33)
14
q =0.1,
15
q =2800,
16
q =0.1,
17
q =0.1,
24
q =0.1,
25
q =2800,
26
q =0.1,
27
q =0.1,
3
r =
4
r =
5
r =
6
r =1,
7
r =
8
r =
9
r =
10
r =1.
Corresponding matrix are
0.4545 0 0 0
0.0011 0.0011 0.0002 0.2043
0 0 0 1
0 0 0 0.0014
A
(
(
(
' =
(
(
,
1
0 0 0 0
0.0021 0 0 0.2032
0 0 0 1
0 0.0004 0.0028 0.0042
B
(
(
(
' =
(
(
,
Applications of Nonlinear Control
72
1 2
0.1 0 0 0
0 2800 0 0
0 0 0.1 0
0 0 0 0.1
C C
(
(
(
' ' = =
(
(
,
12 22
1
1
1
1
D D
(
(
(
' ' = =
(
(
.
Using LMI toolbox, we get state feedback controller:
[ 0.027 697.798 0.023 0.025] F' = (38)
Substituting Eq.(37) and Eq.(38) into Eq.(36), we get nonlinear robust synthetic controller of
dieselgenerator set.
Optimization of weighting function is difficult point of Htwo/Hinfinity control, needing
select repeatedly. After each selection, using LMI toolbox to get state feedback coefficient,
substituting simulation model, making charactrsitc test in order to get best combination
property index. Fig.8 give the block diagram of the system simulation.
Fig. 8. Block diagram of the system simulation
In order to test and verify the effect of nonlinear robust synthetic controller for restraining
the chaotic oscillation of ship power system, Fig.9 and Fig.10 give the phase diagram of
power angle, speed and voltage of two dieselgenerator sets on parallel running with 12.5%
load separately after using nonlinear robust synthetic controller. The initial va1ues of two
sets are same as that of generating chaotic oscillation, the running time is also 100 seconds.
Calculating the Lyapunov index of two dieselgenerator sets on parallel operation, we get
the following results. Lyapunov index of No.1 dieselgenerator set is:
1
0.003435 = ,
2
0.104389 = ,
3
0.230524 = ; Lyapunov index of No.2 dieselgenerator set is:
1
0.003442 = ,
2
0.104382 = ,
3
0.230524 = .
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
73
Fig.11 and Fig.12 give the phase diagram of power angle, speed and voltage of two diesel
generator sets on parallel running with 25% load separately after using nonlinear robust
synthetic controller. The initial va1ues of two sets are same as that of generating chaotic
oscillation, the running time is also 100 seconds. Calculating the Lyapunov index of two
dieselgenerator sets on parallel operation, we get the following results. Lyapunov index of
No.1 dieselgenerator set is:
1
0.003983 = ,
2
0.103822 = ,
3
0.230553 = ; Lyapunov index
of No.2 dieselgenerator set is:
1
0.003993 = ,
2
0.103811 = ,
3
0.230554 = .
Fig.13 and Fig.14 give the phase diagram of power angle, speed and voltage of two diesel
generator sets on parallel running with 25% load plus periodicity load separately after using
nonlinear robust synthetic controller. The initial va1ues of two sets are same as that of
generating chaotic oscillation, the running time is also 100 seconds. Calculating the
Lyapunov index of two dieselgenerator sets on parallel operation, we get the following
results. Lyapunov index of No.1 dieselgenerator set is:
1
0.004331 = ,
2
0.103293 = ,
3
0.230741 = ; Lyapunov index of No.2 dieselgenerator set is:
1
0.004341 = ,
2
0.103283 = ,
3
0.230742 = .
From Fig.9 to Fig.12 we can see, all maximum Lyapunov indexes of the system are less than
0, showing that the system doesnt exists chaotic phenomenon and works in a stable range.
Two dieselgenerator sets with light load on parallel connection, run after using nonlinear
robust synthetic controller, their chaotic phenomenon disappears, power angle, speed and
voltage run nearby the desired values. It shows that nonlinear robust synthetic controller
can control the nonlinear character of ship power system effectively and make the nonlinear
characteristic of ship power system offset and compensate, it can solve the problem of
system chaotic oscillation fundamentally.
0.2
0.25
0.3
0.35
0.4 0.99
0.995
1
1.005
0.9996
0.9996
0.9997
0.9997
0.9998
1
1
U
1
Fig. 9. Phase diagram of No.1 dieselgenerator set when two sets load 12.5% on parallel
connection after using nonlinear robust synthetic controller
Applications of Nonlinear Control
74
0.2
0.25
0.3
0.35
0.4 0.99
0.995
1
1.005
0.9996
0.9996
0.9997
0.9997
0.9998
2
2
U
2
Fig. 10. Phase diagram of No.2 dieselgenerator set when two sets load 12.5% on parallel
connection after using nonlinear robust synthetic controller
0.1
0.2
0.3
0.4
0.5 0.99
0.995
1
1.005
0.9996
0.9997
0.9998
0.9999
1
1
U
1
Fig. 11. Phase diagram of No.1 dieselgenerator set when two sets load 25% on parallel
connection after using nonlinear robust synthetic controller
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
75
0.1
0.2
0.3
0.4
0.5 0.99
0.995
1
1.005
0.9996
0.9997
0.9998
0.9999
2
2
U
2
Fig. 12. Phase diagram of No.2 dieselgenerator set when two sets load 25% on parallel
connection after using nonlinear robust synthetic controller
From Fig.13 to Fig.14 we can see, all maximum Lyapunov indexes of the system are also less
than 0, showing that the system also doesnt exists chaotic phenomenon and works in a
stable range. Although after adding periodicity load intensifying the nonlinearity of system,
nonlinear robust synthetic controller can restrain the chaotic phenomenon of ship power
0.1
0.2
0.3
0.4
0.5 0.98
0.99
1
1.01
1.02
0.998
0.999
1
1.001
1
1
U
1
Fig. 13. Phase diagram of No.1 dieselgenerator set when two sets load 25% plus periodicity
load on parallel connection after using nonlinear robust synthetic controller
Applications of Nonlinear Control
76
0.1
0.2
0.3
0.4
0.5 0.98
0.99
1
1.01
1.02
0.998
0.999
1
1.001
2
2
U
2
Fig. 14. Phase diagram of No.2 dieselgenerator set when two sets load 25% plus periodicity
load on parallel connection after using nonlinear robust synthetic controller
system, making the nonlinear characteristic of ship power system offset and compensate,
thus improving the stability of ship power system.
Because ship power system made up of dieselgenerator sets is a nonlinear system, using
nonlinear robust synthetic controller can offset and compensate the nonlinear characteristic
of ship power system, it can average the load in parallel operation control, there is no power
angular difference between the dieselgenerator sets, thus avoiding the power transmission
between the sets, solving the problem of system chaotic oscillation fundamentally,
improving the voltage and frequency stability of ship power system. The research on
nonlinear robust synthetic controller of dieselgenerator set provides a new control method
for ship power system, having important practical significance and extensive using
prospect.
6. Conclusion
In order to analyze the chaos phenomenon of ship power system, the nonlinear
mathematical model of two dieselgenerator sets on parallel connection is built in this paper,
which reflects the relationship of interaction and mutual influence between two sets. The
light load working condition of two dieselgenerator sets on parallel connection in ship
power station is analyzed by using Lyapunov index method, which proves the presence of
chaos phenomenon. A nonlinear robust synthetic controller is designed which is based on
the nonlinear mathematical model of dieselgenerator set. In combining the direct feedback
linearization with robust control theory to design the synthetic controller for the diesel
generator set, then a nonlinear robust synthetic control law is developed for the diesel
generator set. The computer simulation results show that the nonlinear robust synthetic
Robust Control Research of Chaos
Phenomenon for DieselGenerator Set on Parallel Connection
77
controller effectively suppresses the chaos phenomenon of ship power system, thus
providing desirable stability for ship power system.
7. References
[1] Huang Manlei, Tang Jiaheng, Guo ZhenmingThe Mathematical Model of Diesel
Engine Speed Regulation System[J]Journal of Harbin Engineering University,
1997, 18(6):20~25
[2] Huang Manlei, Li Dianpu, Liu HongdaSimulation Research on Doublepulse Speed
Governor of Diesel Engine[J]SHIP ENGINEERING, 2002, 24(3):36~38
[3] Huang Manlei, Wang ChanghongNonlinear mathematical model of dieselgenerator
set on ship[J]Journal of Harbin Engineering University, 2006, 27(1):1519, 47.
[4] Yang Zhengling, Lin KongyuanStudy on the relation between classical swing
equations and chaos[J]Automation of Electric Power Systems, 2000,
24(7):20~22,45
[5] Jia Hongjie,Yu Yixin,Wang ChengshanChaotic phenomena in power systems and its
studies [J]Proceedings of the CSEE, 2001, 21(7):26~30
[6] Wang Baohua, Yang Chengwu, Zhang Qiang. Summary of bifurcation and chaos
research in electric power system[J]Transactions of China Electrotechnical
Society, 2005, 20(7):1~10
[7] Bernstein D S, Hadded W MLQG control with an H
performance bound: A Riccati
equation approachIEEE Trans on Automatic Control, 1989, 34(3):293305
[8] Zhou K, Glover K, Bodenheimer B, et alMixed H
2
and H
performance objectives I:
Robust performance analysisIEEE Trans on Automatic Control, 1994, 39(8):1564
1574
[9] Doyle J, Zhou K, Glover K, et alMixed H
2
and H
performance objectives II: Optimal
ControlIEEE Trans on Automatic Control, 1994, 39(8):15751587
[10] Khargonekar P P, Rotea M AMixed H
2
/H
control: A convex optimization
approachIEEE Trans on Automatic Control, 1991, 36(7):824837
[11] Sun Yusong, Sun Yuanzhang, Lu Qiang, Shao YixiangResearch on Nonlinear
Robust Control Strategy for Hydroelectric Generators Valve[J]Proceedings of the
CSEE, 2001,21(2):5659,65
[12] Li Wenlei, Jing Yuanwei, Liu XiaopingNonlinear robust control for turbine main
steam valve[J]Control Theory and Applications, 2003,20(3):387390
[13] Wang JinhuaDesign of mixed H
2
/H
n
a + bu + d (1)
where y(t) R is the system output, u(t) R the input, a a vector of varying entries,
n
a known vector, b the control gain, and d a disturbancelike term. We make the following
assumptions:
Ai) The vector a involves unknown, time varying, bounded entries a
1
, , a
j
, which satisfy:
a
1

1
, , a
j

j
, where
1
, ,
j
are unknown, positive constants.
Aii) The entries of the vector
n
are known linear or nonlinear functions of y, , y
(n1)
.
Aiii) The terms y , y, , y
(n1)
are available for measurement.
Aiv) The term d represents either external disturbances or unknown model terms that satisfy:
d
d
f
d
, (2)
where
d
is an unknown positive constant, and f
d
is a known function that depends on y, ,
y
(n1)
. In the case that d is bounded, we have f
d
= 1. The term d may come from the product
of a known function g
d
with an unknown varying but bounded coefcient c
g
: d = c
g
g
d
,
c
g

d
, so that f
d
= g
d
, where
d
is an unknown positive constant whereas g
d
is a known
function.
Av) The control gain b satises:
b b
m
> 0, b = 0 t t
o
(3)
where b
m
is an unknown lower bound, and the value of the signum of b is constant and
known.
Remark. We recall that
d
, b
m
,
1
, ,
j
are unknown constants. In contrast, the values of y,
, y
(n1)
,
n
, f
d
, sgn(b) are required to be known. Notice in assumption Av that we do not
require the control gain b to be upper bounded by any constant. That is a major contribution
with respect to current works that use the Nussbaum gain method, e.g (Su et al., 2009), (Feng,
Hong, Chen & Su, 2008), (Feng, Su & Hong, 2008), (Feng et al., 2007), (Ge & Wang, 2003). The
requirement about the value of the signum of b is a common and acceptable requirement.
3. Control goal
Let
e(t) = y(t) y
d
(t) (4)
y
(n)
d
+ a
m,n1
y
(n1)
d
+ + a
m,o
y
d
= a
m,o
r (5)
e
= {e : e C
be
} (6)
82 Applications of Nonlinear Control
A Robust State Feedback Adaptive Controller with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain 5
where e(t) is the tracking error, y
d
(t) is the desired output,
e
is a residual set, r is the reference
signal. Moreover, a
m,n1
, , a
m,o
are constant coefcients dened by the user, such that
the polynomial K(p) is Hurwitz, being K(p) dened as K(p) = p
(n)
+ a
m,n1
p
(n1)
+ +
a
m,o
. The reference signal r(t) is bounded and userdened. The constant C
be
is positive and
userdened.
The objective of the MRAC design is to formulate a controller, provided by the plant model
(1) subject to assumptions Ai to Av, such that:
i) The tracking error e converges asymptotically to the residual set
e
.
ii) The control signals are bounded and do not involve discontinuous signals.
4. State transformation based on the state backstepping
In this section we carry out a state transformation by following the steps 0, , n of the
backstepping procedure. The plant model (1) can be rewritten as follows:
x
i
= x
i+1
, 1 i n 1 (7)
x
n
= a
n
(x
1
, , x
n
) + bu + d (8)
x
1
= y, x
2
= y, , x
n
= y
(n1)
The model (7, 8) can be obtained by making
1
= =
n1
= 0 in the parametric  pure
feedback form of (Kanellakopoulos et al., 1991). We use the SAB of (Kanellakopoulos et al.,
1991) as the basic framework for the formulation of the control and update laws.
We develop the SAB for the plant model (7, 8), and introduce a new robustness technique.
Since the order of the plant is n, the procedure comprises the steps 0, , n, to be carried out
in a sequential manner.
Step 0. We begin by dening the state z
1
as the tracking error:
z
1
= e = y y
d
= x
1
y
d
(9)
Step i (1 i n 1). At each ith step, we obtain the dynamics of the state z
i
by deriving it
with respect to time, and using the denitions of x
i+1
provided by (7). For the sake of clarity,
we develop the step 1 and then we state a generalization for (1 i n 1).
For the case i = 1, we differentiate z
1
dened in (9) and use the denition of x
1
provided by
(7) with i = 1:
z
1
= x
1
y
d
= x
2
y
d
= x
2
+
1
(10)
1
=
1
( y
d
) = y
d
where
1
is a known function of y
d
. Equation (10) can be rewritten as:
z
1
= c
1
z
1
+ z
2
(11)
z
2
= x
2
+ c
1
z
1
y
d
(12)
where c
1
2 is a positive constant of the user choice. The dynamic equation of z
2
is obtained
by differentiating it with respect to time. The same procedure must be followed until the step
83
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
6 WillbesetbyINTECH
i = n 1. To state a generalization, we express z
i
as:
z
i
= x
i+1
+
i
, (13)
i
=
i
(z
1
, z
i
, y
d
, y
d
, , y
(i)
d
), (14)
where
i
is a known scalar term, that is function of z
1
, , z
i
, y
d
, y
d
, , y
(i)
d
. Equation (13)
can be rewritten as:
z
i
= c
i
z
i
+ z
i+1
, (15)
z
i+1
= x
i+1
+
i
+ c
i
z
i
(16)
where c
i
2 is a positive constant of the user choice. At the step i = n 1 we obtain the
dynamic equation for z
n1
and the expression for z
n
as indicated by (15, 16).
Remark. Notice that the denition of the states z
i
is similar to that of the disturbance free case, so that
a simple design is preserved. This is due to the following facts:
i) Disturbance like terms are absent in the dynamics x
1
, , x
n1
given by (7), so that they are also
absent in the dynamics z
1
, , z
n1
, as can be noticed in (13).
ii) Dead zone functions of the states z
i
are not used.
Step n. We obtain the dynamics of z
n
by differentiating it with respect to time and using the
expression of x
n
provided by (8):
z
n
= bu + a
n
+
n
+ d (17)
n
=
n
(z
1
, , z
n
, y
d
, , y
(n)
d
) (18)
where
n
is a known scalar that is function of z
1
, , z
n
, y
d
, , y
(n)
d
. Notice that the
disturbance like term d and the control input u appear explicitly at the dynamics of z
n
, at
the step n of the procedure. Thus, we have completed the state transformation, which allows
us to develop the controller.
5. Control and update laws
In this section we develop the control and update laws, taking into account the assumptions
stated in section 2 and the goals of section 3. The key elements of the procedure are:
i) Incorporate the assumptions Ai and Aiv, concerning the unknown time varying parameter
a
1
, , a
j
and the disturbance like term d.
ii) Carry out a linear parameterization.
iii) Express the parameterization in terms of adjustment error and adjusted parameter vector.
iv) Introduce the control gain b within the adjusted parameter vector.
v) Formulate the control law.
vi) Formulate a Lyapunovlike function and nd its time derivative.
84 Applications of Nonlinear Control
A Robust State Feedback Adaptive Controller with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain 7
vii) Formulate the update law.
We begin by rewriting (17) as follows:
z
n
= c
n
z
n
+ bu + a
n
+
n
+ c
n
z
n
+ d, (19)
where c
n
2 is a positive constant of the user choice. Multiplying (19) by z
n
, we obtain:
z
n
z
n
= c
n
z
2
n
+ bz
n
u + z
n
a
n
+ z
n
(
n
+ c
n
z
n
) + z
n
d (20)
The term z
n
a
n
+ z
n
(
n
+ c
n
z
n
) + z
n
d can be rewritten as follows:
z
n
a
n
+ z
n
(
n
+ c
n
z
n
) + z
n
d = z
n
(a
n
+ d +
n
+ c
n
z
n
)
= z
n
(a
[1]
n[1]
+ + a
[j]
n[j]
+ d +
n
+ c
n
z
n
)
z
n

_
a
[1]
n[1]
 + +a
[j]
n[j]
 +d +
n
+ c
n
z
n

_
(21)
using assumptions Ai and Aiv of section 2 and parameterizing, we obtain:
z
n
a
n
+ z
n
(
n
+ c
n
z
n
) + z
n
d
z
n

_
1

n[1]
 + +
j

n[j]
 +
d
f
d
+
n
+ c
n
z
n

_
=
b
mn
z
n

(22)
where
=
_

n[1]
, , 
n[j]
, f
d
, 
n
+ c
n
z
n

_
(23)
= (1/
_
b
mn
)
_
1
, ,
j
,
d
, 1
_
(24)
Notice that the entries of the vector are unknown, positive, constant, because bounds of
the time varying parameters a
i
and d have been introduced, according to the properties in
assumptions Ai and Aiv of section 2. Now, we express (22) in terms of adjustment error and
adjusted parameter vector:
z
n
a
n
+ z
n
(
n
+ c
n
z
n
) + z
n
d
_
b
mn
z
n

+
_
b
mn
z
n

(25)
where
=
=
1
b
mn
_
1
, ,
j
,
d
, 1
_
(26)
being
an adjusted parameter vector and
an adjustment error. Using the property (3) in the
term
b
mn
z
n

of (25), we obtain:
_
b
mn
z
n


_
3C
bvz
2
2
3C
bvz
_
bz
n

 (27)
where C
bvz
= (1/2)C
2
be
(28)
85
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
8 WillbesetbyINTECH
using the Youngs inequality (cf. (Royden, 1988) pp. 123), we obtain:
_
b
mn
z
n


3
4
C
bvz
+
1
3C
bvz
bz
2
n
(
)
2
(29)
Substituting (29) into (25), we obtain:
z
n
a
n
+ z
n
(
n
+ c
n
z
n
) + z
n
d
b
mn
z
n

+
3
4
C
bvz
+
1
3C
bvz
bz
2
n
(
)
2
(30)
Remark. We have proposed a new method to handle the unknown varying behavior of the control gain
b, alternative to the current Nussbaum gain method. We parameterized the model term z
n
a
n
+
z
n
(
n
+ c
n
z
n
) + z
n
d in terms of adjustment error
and adjusted parameter vector
, and developed
the following steps:
i) Introduce the constant
b
mn
in the parameterization, see (22).
ii) Introduce the inequality
b
mn
_
b, see (27).
iii) Apply the Youngs inequality to obtain b, see (29).
Recall that the value of b
mn
is not required to be known.
Substituting (30) into (20), we obtain:
z
n
z
n
c
n
z
2
n
+ (3/4)C
bvz
+ bz
n
_
u +
1
3C
bvz
sgn(b)z
n
(
)
2
_
b
mn
z
n

(31)
we choose the following control law:
u =
1
3C
bvz
sgn(b)z
n
(
)
2
(32)
where , z
n
are dened in (23), (16), respectively. Substituting (32) into (31), we obtain:
z
n
z
n
c
n
z
2
n
+ (3/4)C
bvz
_
b
mn
z
n

(33)
To handle the effect of the constant (3/4)C
bvz
, we formulate the following Lyapunovlike
function:
V
z
=
(1/2)(
V
z
C
bvz
)
2
if V
z
C
bvz
0 otherwise
(34)
V
z
= (1/2)(z
2
1
+ + z
2
n
) (35)
where C
bvz
is dened in (28). We need the following properties:
86 Applications of Nonlinear Control
A Robust State Feedback Adaptive Controller with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain 9
Proposition 5.1. The function
V
z
dened in (34) has the following properties:
i)
V
z
0 (36)
ii)V
z
3C
bvz
+3
V
z
(37)
iii)
V
z
, (
V
z
/V
z
) are continuous with respect to V
z
(38)
Proof. From (34) it follows that
V
z
0t t
o
, the property i of proposition 5.1. In addition,
from (34) it follows that
V
z
(
_
2
V
z
+
_
C
bvz
)
2
(39)
Applying the Youngs inequality (cf. (Royden, 1988) pp. 123), we obtain:
V
z
= C
bvz
+2
_
C
bvz
_
2
V
z
+2
V
z
3C
bvz
+3
V
z
(40)
This completes the proof of property ii. From (42) it follows that
V
z
/V
z
= 0 if V
z
= C
bvz
and that
V
z
/V
z
is continuous. From (34) it follows that
V
z
is continuous. This completes the
proof of property iii of proposition 5.1.
Differentiating (34) with respect to time, we obtain:
d
V
z
dt
=
V
z
V
z
V
z
(41)
V
z
V
z
=
_
(1/2)(1/
V
z
)(
V
z
C
bvz
) if V
z
C
bvz
0 otherwise
(42)
To compute
V
z
, we differentiate (35) with respect to time:
V
z
= z
1
z
1
+ + z
n
z
n
. Introducing
(11) and (15), we obtain
V
z
= c
1
z
2
1
+ z
1
z
2
c
2
z
2
2
+ z
2
z
3
+ + z
n
z
n
(43)
substituting (33), we obtain:
V
z
c
1
z
2
1
+ z
1
z
2
c
2
z
2
2
+ z
2
z
3
+ c
n
z
2
n
+(3/4)C
bvz
b
mn
z
n

(44)
Provided that c
1
2, c
2
2, , c
n
2 and completing the squares yields:
c
1
z
2
1
+ z
1
z
2
c
2
z
2
2
+ z
2
z
3
c
n
z
2
n
z
2
1
(3/4)z
2
2
+ (3/4)z
2
n
(3/2)V
z
substituting into (44), we obtain:
V
z
(3/2)V
z
+ (3/4)C
bvz
b
mn
z
n

(45)
87
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
10 WillbesetbyINTECH
Since
V
z
/V
z
is nonnegative, we can multiply it into (45) without changing the direction of
the inequality:
V
z
V
z
V
z
(3/2)V
z
V
z
V
z
+ (3/4)C
bvz
V
z
V
z
_
b
mn
z
n

V
z
V
z
(46)
Substituting into (41), we obtain:
d
V
z
dt
(3/2)V
z
V
z
V
z
+ (3/4)C
bvz
V
z
V
z
_
b
mn
z
n

V
z
V
z
(47)
we choose the update law so as to reject the effect of the term involving the adjustment error
= z
n

V
z
V
z
(48)
where is a diagonal matrix whose diagonal elements are positive constants dened by the
user, whereas , z
n
,
V
z
/V
z
are dened in (23), (16), (42), respectively.
Remark. So far, we have developed the controller, which involves the control law (32) and the update
law (48). Other parameters necessary for its implementation are: V
z
dened in (35); z
1
, z
2
, , z
n
dened in (9), (12), , (16), respectively; C
bvz
dened in (28). Recall that c
1
2, , c
n
2 are
userdened positive constants.
Remark. The control and update laws stated in (32) and (48) have the following features:
i) The control law uses the adjusted parameter vector
, so that it does not rely on upper or lower bounds
of the plant coefcients, i.e.
1
, ,
j
,
d
, b
mn
, and excessive control effort is also avoided.
ii) Additional states are not required to handle the unknown varying behavior of the control gain, what
is an important benet with respect to closely related schemes that use the Nussbaum gain method.
iii) The control and update laws do not involve discontinuous signals. In fact, the vectorial eld of the
closed loop system is Lipschitz continuous, so that trajectory unicity is preserved.
6. Boundedness analysis
In this section we prove that the closed loop signals z
1
, , z
n
,
, u are bounded if the
developed controller is applied.
Theorem 6.1. Boundedness of the closed loop signals. Consider the plant (1) subject to
assumptions Ai to Av; the signals z
1
, , z
n
dened in (9), (12) and (16); the signals , V
z
,
V
z
/V
z
,
C
bvz
dened in (23), (35), (42) and (28), respectively. If the controller (32), (48) is applied, then the
signals z
1
, , z
n
,
, and u remain bounded and e is upper bounded as follows:
e
2
_
_
C
bvz
+
_
2V( x(t
o
))
_
2
(49)
Proof. We choose the following Lyapunovlike function:
V( x(t)) =
V
z
+ V
(50)
88 Applications of Nonlinear Control
A Robust State Feedback Adaptive Controller with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain 11
V
= (1/2)
_
b
mn
(51)
x(t) = [z
1
, . . . , z
n
,
(52)
where
V
z
is dened in (34) and
in (26). The time derivative of V
is:
= (1/2)
_
b
mn
(
) (53)
Since is diagonal, then
1
is diagonal, (
1
)
=
1
,
. In view of this
and the update law (48), we have:
=
_
b
mn
=
_
b
mn
z
n

V
z
V
z
(54)
Differentiating (50) with respect to time, we obtain:
V =
V
z
+
V
V (3/2)V
z
V
z
V
z
+ (3/4)C
bvz
V
z
V
z
=
3
2
V
z
V
z
_
V
z
2
+
V
z
2
C
bvz
2
_
(55)
From (42) if follows that
V
z
/V
z
= 0 for V
z
C
bvz
(56)
V
z
/V
z
> 0 for V
z
> C
bvz
. (57)
In view of this and (55), we obtain:
V
3
2
V
z
V
z
_
V
z
2
_
if V
z
C
bvz
V 0 =
3
2
V
z
V
z
_
V
z
2
_
if V
z
< C
bvz
V
3
4
V
z
V
z
V
z
0 (58)
Thus,
V +0V 0. Using the Lemma in (Slotine & Li, 1991) pp. 91, we obtain:
V( x(t)) V( x(t
o
)) exp(0t) = V( x(t
o
)) (59)
where
V( x(t
o
)) =
V
zo
+ V
o
(60)
V
zo
=
_
(1/2)(
V
zo
C
bvz
)
2
if V
zo
C
bvz
0 otherwise
(61)
V
zo
= (1/2)(z
1
(t
o
)
2
+ + z
n
(t
o
)
2
) (62)
V
o
= (1/2)
_
b
mn
(
(t
o
) )
1
(
(t
o
) ) (63)
89
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
12 WillbesetbyINTECH
Since V( x(t)) 0, we have: 0 V( x(t)) V( x(t
o
)). Introducing the denition (50), we
obtain
V
z
+ V
V( x(t
o
)) (64)
V
z
V( x(t
o
)), V
V( x(t
o
)) (65)
Thus, it follows from(51) that
L
, and consequently
L
. The inequality
V
z
V( x(t
o
))
implies that the tracking error e is bounded, as we show hereafter. We begin by solving (34)
for V
z
:
V
z
=
_
_
C
bvz
+
_
2
V
z
_
2
if
V
z
> 0 (66)
V
z
C
bvz
if
V
z
= 0 (67)
Using the inequality
V
z
V( x(t
o
)), we obtain:
V
z
_
_
C
bvz
+
_
2V( x(t
o
))
_
2
if
V
z
> 0 (68)
V
z
C
bvz
_
_
C
bvz
+
_
2V( x(t
o
))
_
2
if
V
z
= 0 (69)
combining both inequalities, we obtain:
V
z
_
_
C
bvz
+
_
2V( x(t
o
))
_
2
(70)
Introducing the denition (35), we obtain:
_
z
2
1
+ + z
2
n
2
_
_
C
bvz
+
_
2V( x(t
o
))
_
2
(71)
so that z
1
L
, , z
n
L
. Since e
2
= z
2
1
z
2
1
+ + z
2
n
, we obtain:
e
2
_
_
C
bvz
+
_
2V( x(t
o
))
_
2
(72)
which indicates the upper bound for the tracking error e.
Remark. Notice that this upper bound does not involve integral terms, what is an important advantage
with respect to the NussbaumGain method, see (Su et al., 2009), (Feng, Hong, Chen &Su, 2008), (Feng
et al., 2006), (Ge & Wang, 2003), (Feng et al., 2007), (Feng, Su & Hong, 2008), (Ren et al., 2008).
Now, we proceed to show the boundedness of u. From (9), (12), (16) it follows that x
1
L
,
x
2
L
, , x
n
L
. Therefore,
n
L
. From (32) it
follows that u L
V f
d
0 (73)
f
d
=
_
(3/8)(
V
z
)(
V
z
C
bvz
) if V
z
C
bvz
0 otherwise
(74)
The derivative f
d
/V
z
is not continuous, as it involves an abrupt change at V = C
bvz
. Thus,
the Barbalats Lemma can not be applied on f
d
. To remedy that, we shall express (73) in terms
of a function with continuous derivative:
V f
d
f
g
0 (75)
f
g
=
_
(3/8)(
V
z
C
bvz
)
2
if V
z
C
bvz
0 otherwise
(76)
Arranging and integrating (75), we obtain:
f
g
V
_
t
to
f
g
d V( x(t
o
)) V( x(t))
V( x(t)) +
_
t
to
f
g
d V( x(t
o
)) (77)
Thus, f
g
L
1
. We have to prove that f
g
L
,
f
g
L
f
g
=
f
g
V
z
V
z
(78)
f
g
V
z
=
_
(3/8)(1/
V
z
)(
V
z
C
bvz
) if V
z
C
bvz
0 otherwise
(79)
Notice that f
g
/V
z
is continuous. Since V
z
L
, then f
g
/V
z
L
. Since z
1
L
, ,
z
n
L
, , z
n1
L
. Since u L
, it follows from
(17) that z
n
L
.
So far we have proved that f
g
/V
z
L
and
V
z
L
f
g
L
. In view of f
g
L
,
f
g
L
2
a + bu + d (80)
2
= [ y, y]
, a = [a
1
, a
2
]
(81)
a
1
= 2 (1 +0.1sin(2(/8)t)) , a
2
= 1 (1 +0.1sin(2(/5)t))
b = 2 (1 +0.1sin((2/11)t)) +0.6y
d = 0.2 (1 +0.1sin((2/7)t)) y
(82)
The aim is that y converges towards y
d
, with a threshold of 0.1. In gure 1 we present a
simulation block diagram for the example.
Fig. 1. Simulation block diagram.
The properties Ai, Aiv, Av of section 2 are analyzed at the following. From (82) it follows that
a
1
, d, b are bounded as:
a
1
 2(1.1) = 2.2, a
2
 1(1.1) = 1.1 (83)
d 0.2(1.1)y = 0.22y (84)
b 2(0.9) = 1.8 > 0 (85)
Hence, the upper bounds of a
1
, a
2
, d, and the lower bound of b, are:
a
1

1
= 2.2, a
2

2
= 1.1, b b
mn
= 1.8 (86)
d
d
f
d
,
d
= 0.22, f
d
= y (87)
where f
d
is not constant and known, whereas b
mn
,
1
,
2
,
d
, b
mn
are positive, constant and
unknown to the controller. From (86), (87) it follows that assumptions Ai, Aiv, Av of section 2
are satised.
The procedure of section 4 is followed in order to establish the terms involved in the control
and update laws, mentioned in remark 5. Eq. (80) can be rewritten as:
x
1
= x
2
(88)
x
2
=
2
a + bu + d (89)
x
1
= y, x
2
= y, n = 2 (90)
92 Applications of Nonlinear Control
A Robust State Feedback Adaptive Controller with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain 15
since n = 2, the state transformation based on the backstepping procedure involves the steps
0, 1, 2.
Step 0. Let
z
1
= e = y y
d
= x
1
y
d
(91)
as in (9).
Step 1. Differentiating (91) with respect to time and arranging, yields:
z
1
= c
1
z
1
+ z
2
(92)
z
2
= x
2
+ c
1
z
1
y
d
(93)
as in (11), (12).
Step 2. Since n = 2, the second step is the last one. Differentiating (93) with respect to time,
using (89) and arranging, yields:
z
2
= x
2
+ c
1
z
1
y
d
=
2
a + bu + d + c
1
z
1
y
d
=
2
a + bu + d + c
1
(x
2
+
1
) y
d
(94)
using the denitions (91), (93), yields:
z
2
=
2
a + bu + d +
2
(95)
2
= c
1
(z
2
c
1
z
1
) y
d
(96)
notice that the form of (95), (96) is that of (17), (18), respectively. This completes the state
transformation based on the backstepping procedure.
The parameters dened above can be summarized as:
z
1
= y y
d
(97)
z
2
= x
2
+ c
1
z
1
y
d
(98)
x
1
= y, x
2
= y
d
(99)
1
= y
d
(100)
2
= c
1
(z
2
c
1
z
1
) y
d
(101)
According to remark 5, it remains to dene , V
z
. From (81), denition (23) and n = 2, it
follows that
=
_

2[1]
, 
2[2]
, f
d
, 
2
+ c
2
z
2

_
= [ y, y, f
d
, 
2
+ c
2
z
2
]
(102)
From (35) and n = 2 it follows that
V
z
= (1/2)(z
2
1
+ z
2
2
) (103)
93
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
16 WillbesetbyINTECH
Expressions (97) to (103) allow to dene the control and update law. From (32), (48), (82) and
n = 2 it follows that
sgn(b) = +1 (104)
u =
1
3C
bvz
z
2
(
)
2
(105)
= z
2

V
z
V
z
(106)
the main parameters needed to compute u and
are: (102),
2
(101), C
bvz
(28), z
2
(98), z
1
(97),
V
z
/V
z
(42), V
z
(103). In addition, is a diagonal matrix whose diagonal elements are
positive constants dened by the user.
0 20 40 60 80 100
0
1
2
3
time
y
,
y
d
0 20 40 60 80 100
0.1
0
0.1
time
e
0 20 40 60 80 100
0
2
4
time
u
Fig. 2. Example 1, upper: output y (continuous line), desired output y
d
(dashdot line);
middle: tracking error e; lower: control input u.
Since the aim is that y converges towards y
d
, with a threshold of 0.1, we set C
be
= 0.1.
We use the reference model (5) with y
d
(t
o
) = y(t
o
), y
d
(t
o
) = 0, a
m,1
= 1, a
m,o
= 1.
We use the following parameter values for the control and update laws: c
1
= 2, c
2
= 2,
= diag{1, 1, 1, 1}.
The results are shown in gures 2 and 3. We have choosen y
d
(t
0
) y(t
0
) in order to obtain a
rapid convergence of y towards y
d
. Figure 2 shows that. i) the tracking error e converges
asymptotically towards
e
= {e : e 0.1}. ii) The output y converges towards y
d
with threshold 0.1 without large transient differences. Figure 3 shows that
1
, ...,
4
are not
decreasing with respect to time. This occurs because
1
0 20 40 60 80 100
0
0.1
0.2
time
^
2
0 20 40 60 80 100
0
0.1
0.2
time
^
3
0 20 40 60 80 100
0
0.2
0.4
time
^
4
Fig. 3. Example 1, entries of the updated parameter vector
, from upper to lower:
1
;
2
,
3
,
4
.
9. Acknowledgements
A. Rincon acknowledges nancial support provided by Programa de becas para estudiantes
sobresalientes de posgrado, Universidad Nacional de Colombia  vicerrectora de
Investigacin. This work was partially supported by Universidad Nacional de Colombia 
Manizales, project 12475, Vicerrectora de Investigacin, DIMA.
10. References
Bashash, S. & Jalili, N. (2009). Robust adaptive control of coupled parallel piezoexural
nanopositioning stages, IEEE/ASME Transactions on Mechatronics Vol. 14(No.
1): 1120.
Bechlioulis, C. & Rovithakis, G. (2009). Adaptive control with guaranteed transient and
steady state tracking error bounds for strict feedback systems, Automatica Vol. 45(No.
2): 532538.
Cao, C. & Hovakimyan, N. (2006). Design and analysis of a novel l
1
adaptive controller,
part i: control signal and asymptotic stability, Proceedings of the 2006 American Control
Conference, Publisher, Minneapolis (USA), pp. 33973402.
Cao, C. & Hovakimyan, N. (2008a). Design and analysis of a novel adaptive control
architecture with guaranteed transient performance, IEEE Transactions on Automatic
Control Vol. 53(No. 2): 586591.
Cao, C. & Hovakimyan, N. (2008b). l
1
adaptive output feedback controller for systems of
unknown dimension, IEEE Transactions on Automatic Control Vol. 53(No. 3): 815821.
Chen, B., Liu, X., Liu, K. & Lin, C. (2009). Novel adaptive neural control design for nonlinear
mimo timedelay systems, Automatica Vol. 45(No. 6): 15541560.
95
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
18 WillbesetbyINTECH
Chen, C., Lin, C. & Chen, T. (2008). Intelligent adaptive control for mimo uncertain nonlinear
systems, Expert Systems with Applications Vol. 35(No. 3): 865 877.
Chen, W. (2009). Adaptive backstepping dynamic surface control for systems with periodic
disturbances using neural networks, IET Control Theory and Applications Vol. 3(No.
10): 13831394.
Dobrokhodov, V., Kitsios, I., Kaminer, I. &Jones, K. (2008). Flight validation of a metrics driven
l
1
adaptive control, Proceedings of AIAA Guidance, Navigation and Control Conference
and Exhibit, Publisher, Honolulu (USA), pp. 122.
Du, H., Ge, S. & Liu, J. (2010). Adaptive neural network output feedback control for a class
of nonane nonlinear systems with unmodelled dynamics, IET Control Theory and
Applications Vol. 5(No. 3): 465477.
Feng, Y., Hong, H., Chen, X. & Su, C. (2008). Robust adaptive controller design for a
class of nonlinear systems preceded by generalized prandtlishlinskii hysteresis
representation, Proceedings of 7th World Congress on Intelligent Control and Automation,
Publisher, Chongqing (China), pp. 382387.
Feng, Y., Hu, Y. &Su, C. (2006). Robust adaptive control for a class of perturbed strictfeedback
nonlinear systems with unknown prandtlishlinskii hysteresis, Proceedings of the 2006
IEEE International Symposium on Intelligent Control, Publisher, Munich (Germany),
pp. 106111.
Feng, Y., Su, C. & Hong, H. (2008). Universal contruction of robust adaptive control
laws for a class of nonlinear systems preceded by generalized prandtlishlinskii
representation, Proceedings of the 3rd IEEE Conference on Industrial Electronics and
Applications, Publisher, Singapore, pp. 153158.
Feng, Y., Su, C., Hong, H. & Ge, S. (2007). Robust adaptive control for a class of nonlinear
systems with generalized prandtlishlinskii hysteresis, Proceedings of the 46th IEEE
Conference on Decision and Control, Publisher, New Orleans (LA, USA), pp. 48334838.
Ge, S. & Tee, K. (2007). Approximationbased control of nonlinear mimo timedelay systems,
Automatica Vol. 43(No. 1): 3143.
Ge, S. & Wang, J. (2003). Robust adaptive tracking for timevarying uncertain nonlinear
systems with unknown control coefcients, IEEE Transactions on Automatic Control
Vol. 48(No. 8): 14631469.
Ho, H., Wong, Y. & Rad, A. (2009). Adaptive fuzzy sliding mode control with chattering
elimination for nonlinear siso systems, Simulation Modelling Practice and Theory Vol.
17(No. 7): 11991210.
Hong, Y. & Yao, B. (2007). A globally stable saturated desired compensation adaptive robust
control for linear motor systems with comparative experiments, Automatica Vol.
43(No. 10): 18401848.
Hsu, C., Lin, C. &Lee, T. (2006). Wavelet adaptive backstepping control for a class of nonlinear
systems, IEEE Transactions on Neural Networks Vol. 17(No. 5): 11751183.
Huang, A. & Kuo, Y. (2001). Sliding control of nonlinear systems containing timevarying
uncertainties with unknown bounds, International Journal of Control Vol. 74(No.
3): 252264.
Ioannou, P. & Sun, J. (1996). Robust Adaptive Control, PrenticeHall PTR Upper Saddle River,
New Jersey.
96 Applications of Nonlinear Control
A Robust State Feedback Adaptive Controller with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain 19
Jiang, Z. & Hill, D. (1999). A robust adaptive backstepping scheme for nonlinear systems with
unmodeled dynamics, IEEE Transactions on Automatic Control Vol. 44(No. 9): 1705
1711.
Kanellakopoulos, I., Kokotovi c, P. & Morse, A. (1991). Systematic design of adaptive
controllers for feeedback linearizable systems, IEEE Transactions on Automatic Control
Vol. 36(No. 11): 1241 1253.
Koo, K. (2001). Stable adaptive fuzzy controller with timevarying dead zone, Fuzzy Sets and
Systems Vol. 121(No. 1): 161168.
Labiod, S. & Guerra, T. (2007). Adaptive fuzzy control of a class of siso nonafne nonlinear
systems, Fuzzy Sets and Systems Vol. 158(No. 10): 1126 1137.
Li, D. & Hovakimyan, N. (2008). Filter design for feedbackloop tradeoff of l
1
adaptive
controller: a linear matrix inequality approach, Proceedings of IAA Guidance,
Navigation and Control Conference and Exhibit, Publisher, Honolulu (USA), pp. 112.
Li, Y., Qiang, S., Zhuang, X. & Kaynak, O. (2004). Robust and adaptive backstepping control
for nonlinear systems using rbf neural networks, IEEE Transactions on Neural Networks
Vol. 15(No. 3): 693701.
Liu, C. & Tong, S. (2010). Fuzzy adaptive decentralized control for nonlinear systems
with unknown highfrequency gain sign based on klter, Proceedings of Control and
Decision Conference, Publisher, China, pp. 15761581.
Nakanishi, J., Farrell, J. & Schaal, S. (2005). Composite adaptive control with locally weighted
statistical learning, Neural Networks Vol. 18(No. 1): 7190.
Park, B., Yoo, S., Park, J. & Choi, Y. (2009). Adaptive neural sliding mode control of
nonholonomic wheeled mobile robots with model uncertainty, IEEE Transactions on
Control Systems Technology Vol. 17(No. 1): 201214.
Psillakis, H. (2010). Further results on the use of nussbaum gains in adaptive neural network
control, IEEE Transactions on Automatic Control Vol. 55(No. 12): 28412846.
Ren, B., Ge, S., Lee, T. & Su, C. (2008). Adaptive neural control for uncertain nonlinear systems
in purefeedback form with hysteresis input, Proceedings of the 47th IEEE Conference
on Decision and Control, Publisher, Cancun (Mexico), pp. 8691.
Royden, H. (1988). Real Analysis, Prentice Hall Upper Saddle River, New Jersey.
Slotine, J. & Li, W. (1991). Applied Nonlinear Control, Prentice Hall Englewood Cliffs, New
Jersey.
Su, C., Feng, Y., Hong, H. & Chen, X. (2009). Adaptive control of system involving complex
hysteretic nonlinearities: a generalised prandtlishlinskii modelling approach,
International Journal of Control Vol. 82(No. 10): 17861793.
Tong, S., Liu, C. & Li, Y. (2010). Fuzzyadaptive decentralized outputfeedback control for
largescale nonlinear systems with dynnamical uncertainties, IEEE Transaction on
Fuzzy Systems Vol. 18(No. 5): 11.
Tong, S., Tang, J. & Wang., T. (2000). Fuzzy adaptive control of multivariable nonlinear
systems, Fuzzy Sets and Systems Vol. 111(No. 2): 153167.
Wang, X., Su, C. & Hong, H. (2004). Robust adaptive control of a class of nonlinear systems
with unknown deadzone, Automatica Vol. 40(No. 3): 407413.
Wen, C., Zhou, J. & Wang, W. (2009). Decentralized adaptive backst epping stabilization of
interconnected systems with dynamic input and output interactions, Automatica Vol.
45(No. 1): 5567.
97
A Robust State Feedback Adaptive Controller
with Improved Transient Tracking Error Bounds for Plants with Unknown Varying Control Gain
20 WillbesetbyINTECH
Yao, B. (1997). High performance adaptive robust control of nonlinear systems: a general
framework and new schemes, Proceedings of the 36th Conference on Decision and
Control, Publisher, San Diego (California, USA), pp. 24892494.
Yao, B. & Tomizuka, M. (1994). Robust adaptive sliding mode control of manipulators
with guaranteed transient performance, Proceedings of the American Control Conference,
Publisher, Baltimore (Maryland, USA), pp. 11761180.
Yao, B. & Tomizuka, M. (1997). Adaptive robust control of siso nonlinear systems in a
semistrict feedback form, Automatica Vol. 33(No. 5): 893900.
Yousef, H. & Wahba, M. (2009). Adaptive fuzzy mimo control of induction motors, Expert
Systems with Applications Vol. 36(No. 3): 41714175.
Zhang, T. & Ge, S. (2009). Adaptive neural network tracking control of mimo nonlinear
systems with unknown dead zones and control directions, IEEE Transactions on
Neural Networks Vol. 20(No. 3): 483497.
Zhou, J., Wen, C. & Wang, W. (2009). Adaptive backstepping control of uncertain systems
with unknown input timedelay, Automatica Vol. 45(No. 6): 14151422.
Zhou, J., Wen, C. & Zhang, Y. (2004). Adaptive backstepping control of a class of uncertain
nonlinear systems with unknown backlashlike hysteresis, IEEE Transactions on
Automatic Control Vol. 49(No. 10): 17511757.
98 Applications of Nonlinear Control
6
A Robust Motion Tracking Control
of PiezoPositioning Mechanism
with Hysteresis Estimation
Amir Farrokh Payam,
Mohammad Javad Yazdanpanah and Morteza Fathipour
Department of Electrical and Computer Engineering, University of Tehran, Tehran,
Iran
1. Introduction
Piezoelectric actuators are the most suited actuation devices for high precision motion
operations in the positioning tasks include miro/nanopositioning [1]. These actuators have
unlimited motion resolution and posses some advantages such as ignorable friction,
noiseless, zero backlash and easy maintenance, in comparison with the conventional
actuated systems which are based on the sliding or revolute lower pairs [2].
Producing large forces, fast response and high efficiency are major advantages of
piezoelectric actuators. But, it has some drawbacks such as hysteresis behavior, drift in time,
temperature dependence and vibration effects. Molecular friction at sites of materials
imperfections due to domain walls motion is the general cause of hysteresis in piezoelectric
materials [3]. The hysteresis is a major nonlinearity for piezoactuators and often limits
system performance via undesirable oscillations or instability. Therefore, it is difficult to
obtain an accurate trajectory tracking control. Numerous mathematical methods have been
proposed to analyze the hysteresis behavior of piezoelectric actuators. These studies may be
categorized in asymmetrical and symmetrical methods. The asymmetrical types of hysteretic
models include polynomial model [4], Preisachs model [5], neural network model [6] and
Karasnoselskii and Pokrovskii [7]. The symmetrical types of hysteretic model include
Duhem model [8], BoucWen model [9] and Lugre model [9].
The asymmetrical methods establish the nonlinear relations between the input and output
based on the measured input/output data sets. Because superposition of a basic hysteresis
operator is a fundamental principle in these models, they are also called to operator based
model. Although, an operator based model may give a good match with experimental data,
the dynamics of the piezoelectric material is not formulated in these modeling methods and
model parameter identification and implementation is more difficult in this case. The
symmetrical methods employ nonlinear differential equations in order to describe
hysteresis. In this case, the dynamics of the piezoelectric materials are described but the non
symmetric hysteresis is not modeled. However these models are more tractable for control
design. In order to include the hysteresis effect and compensating its effect, Lugre model
[10] is analyzed and studied in this chapter.
Applications of Nonlinear Control
100
There are several control methods to overcome the above mentioned errors and increase the
tracking control precision of the piezoelectric actuators. Some of these methods are PI and
PID controller, fuzzy controller [11], adaptive RFNN [12], feedforward model reference
control method [13], adaptive hysteresis inverse cascade with the plant [14], reinforcement
discrete neuroadaptive controller [15], adaptive wavelet neural network controller [16],
nonlinear observerbased slidingmode controller [17] an adaptive backstepping controller
[10, 18], robust motion tracking controller based on slidingmode theory [19] and continuous
time controller based on SMC and disturbance observer [20]. In some of these works, a
complex inverse hysteresis model has been adopted to overcome the nonlinear hysteresis
effect. Also, in the methods based on the neural network approach, to ensure the error is
bounded, it is assumed that the system states must be inside a compact set. Moreover,
robustness against parameters uncertainties and external disturbances is the other problem
encounter the control methods presented for piezoelectric actuator.
In this chapter, a robust motion tracking control strategy in combination with the hysteresis
force observer is designed and investigated for the piezoelectric actuator. The presented
controller is robust against the unknown or uncertain system parameters and can estimate
the hysteresis force with its estimation property. This control strategy is established based
on the lumped parameter dynamic model. Using Lyapanouv stability analysis, the stability
analysis of the overall control and observer system is performed. Furthermore, the validity
and effectiveness of the designed methodology is investigated by numerical analysis and its
results obtained are compared with those of [19].
Section 2 contains the model explanation of piezopositioning mechanism. A hysteresis
model for piezoelectric systems based on Lugre model is discussed in section 3. Design
procedure of the developed controller is presented in section 4. Simulation analysis and
results obtained are presented in section 5. Finally, section 6 includes the conclusion of this
chapter.
2. Model of piezopositioning mechanism with hysteresis
The lumped parameter dynamic model of the piezodriven mechanism is written as [10, 16]:
H L
Mx Dx F F u + + + = (1)
Where Mis the mass of the controlled piezopositioning mechanism, D is the linear friction
coefficient of the piezodriven system,
L
F denotes the external load,
H
F is the hysteresis
friction force function, ( ), ( ), x t x t and ( ) x t denote the piezoelectric displacement, velocity and
acceleration, respectively and u is the applied voltage to the piezopositioning mechanism.
A block diagram of the model (1) is depicted in Fig.1. Note that the simulation program
used in this chapter is C++.
Noted that:
0
( )
H H
F t F o s ,
1
( )
H H
F t F o s
,
2
( )
H H
F t F o s
,,
0
( )
L L
F t F o s
1
( )
L L
F t F o s
,
2
( )
L L
F t F o s
(2)
Where
1
, 0, 1, 2
Hi
F i o e9 = and
1
, 0, 1, 2
Li
F i o e9 = denote the known upper bounds.
A Robust Motion Tracking Control
of PiezoPositioning Mechanism with Hysteresis Estimation
101
Fig. 1. Model of piezopositioning mechanism with hysteresis [10].
3. Lugre hysteresis model for piezopositioning system
This chapter uses Lugre model to describe nonlinear hysteretic curve of piezoelectric
systems. The mathematical equation is as follows [10,16]:
( )
0 1 1 2
1
( )
H
F z z x x
g x
o o o o = + +
(3)
( )
x
z x z
g x
=
(4)
Where z interpreted as the contact force applied voltage average bristle coefficient,
0 1 2
, , o o o are positive constants and generally can be equivalently interpreted a bristle
stiffness, damping and viscousdamping coefficients, respectively. Moreover, the function
( ) g x denotes the Stribeck effect curve described by:
( )
( )
2
/
0
( )
S
x x
C S C
g x f f f e o
= +
(5)
Where
C
f is the coloumb friction level,
S
f is the level of stiction force and
S
x is the Stribeck
velocity. The function ( ) g x is positive and constant [16]. As depicted in [10], the following
lemma is held.
Lemma 1. Consider nonlinear dynamics system of (4). For any piecewise continuous signal
x and x , the output ( ) z t is bounded.
Substituting the (3) in (1) and arranging the expression, the following dynamics equation is
obtained for the piezopositioning mechanism with Lugre hysteresis model:
Applications of Nonlinear Control
102
( ) ( )
0 1 1 2
1 1
( )
L
z x D
x u z F x x
M M g x M
o o o o
(
= + + +
(
(6)
4. Design controller
In this section we design a robust control strategy for the system of (1), to asymptotically
estimate the hysteresis parameter of equation (3). To facilitate the design process, we
assumed that the hysteresis force is dependent only on the time and
H
F and its first two
time derivatives remain bounded for all times. Using (6) we can rewrite equation (1) as:
u F D
x x
M M M
= (7)
Where ( ) ( )
0 1 1 2
( )
L
z x
F z F x
g x
o o o o = + + +
( ) ( )sgn( )
d
u M x e Dx F x s o  = + + + + (17)
Where
1
(18)
Where sgn(.) is the standard signum function, and
1
1
, k e9 are positive constants.
Noted that the equation (18) for
(19)
Due to the boundness of F and its first two time derivatives, it is understand that
( ), ( ) t t L q q
e .
Based on the stability analysis, the constant should be chosen to satisfy the following
inequality:
1
( ) ( ) t t q q

> + (20)
By substituting (17) in (14), and simplify the obtained expression, the dynamics of r is
achieved as:
1 1
1
( ( ) ( )sgn( )) r F F x x s
M
_  = + (21)
Where
1
( ) x Mx Dx _ = A + A .
Applications of Nonlinear Control
104
Now, using (21), the time derivative of (16) is obtained as:
1 1 1 1
1 1
( ( ) ( )sgn( )) ( ( ( ) ( )sgn( ))) s F F x x s F F x x s
M M
_   _  = + + +
(22)
Substituting (18) and (19) in (22), gives:
1
1 1 1 1 1
1
sgn( ) ( ( ) ( )sgn( )) ( ( ) ( )sgn( ))
k
s k s r x x s x x s
M M

q _  _ 
+
= + +
(23)
Remark 1: If ( ) s t L
e , then ( ), ( ) r t r t L
= (24)
And global asymptotetic estimation of hysteresis in the sense that:
lim[ ] 0
t
F F
= (25)
With the constant satisfies the condition of (20).
Proof
A nonnegative, scalar function
1
( ) V t e9 is defined as:
2
1
2
V s = (26)
Tacking time derivative of (26) gives:
2 1
1 1 1 1 1
( )( sgn( )) ( ( ) ( )sgn( )) ( ( ) ( )sgn( ))
s k
V k s r r r x x s s x x s
M M

 q _  _ 
+
= + + + +
(27)
From (15) it is clear that:
1
1 1 1 1
1
( ( ) ( )sgn( )) ( ( ) ( )sgn( )) 0
k
x x s x x s
M M

_  _ 
+
+ s
(28)
Where
1
( ) x M x D x  o o = +
.
Hence:
2
1
( )( sgn( )) V k s r r r  q s + +
(29)
After integrating both sides of (29), the following inequality is obtained:
A Robust Motion Tracking Control
of PiezoPositioning Mechanism with Hysteresis Estimation
105
0 0
2
0 1 0
1 ( )
( ) ( ) ( ) ( ) ( ) ( ) ( )
t t
t t
d
V t V t k s d r t t r t r d
d
q o
o o q  o q o ,
 o
(  
( s + + + +
( 
\ .
} }
(30)
Where
1
0
, e9 is a positive constant, defined by:
0 0 0 0
( ) ( ) ( ) r t t r t , q = + (31)
After applying the (20) to the bracketed term of (30), ( ) V t can be upper bounded as follows:
0
2
0 1 0
( ) ( ) ( )
t
t
V t V t k s d o o , s +
}
(32)
It is deduced from (32) and (26) that ( ) V t L
e and ( ) s t L
e .
Equation (32) rearrange as:
0
2
1 0 0
( ) ( ) ( )
t
t
k s d V t V t o o , s +
}
(33)
From the fact that ( ) V t is nonnegative and equation (33), it can be deduced that
2
( ) s t L e .
Because of
2
( ) s t L L
e and ( ) s t L
= (34)
Therefore:
lim ( ), ( ), ( ), ( ) 0
t
r t r t e t e t
=
(35)
6. Simulation results
In this section, simulation results are presented to investigate the performance of the
presented method for piezoelectric actuator. First, we test the controller response with the
nominal value of piezoelectric actuator, when 0.3sin( )
L
F t = and initial values are:
(0) 0, (0) 0, (0) 0 x x z = = = . The objective of the positioning is to drive the displacement signal
x to track the reference trajectory which is shown in Fig.2.
The parameters of piezopositioning mechanism are given in Table 1.
0
50000 / N m o =
4
1
5 10 / Ns m o = 2
0.4 / Ns m o =
1
C
f N = 1.5
S
f N = 0.001 /
S
x m s =
1 M kg = 0.0015 / D Ns m = 0.3sin( )
L
F t =
Table 1. Piezoelectric Parameters
Applications of Nonlinear Control
106
0 0.2 0.4 0.6 0.8 1
15
10
5
0
5
10
15
Time (Sec)
D
i
s
p
l
a
c
e
m
e
n
t
(
M
i
c
r
o
m
e
t
e
r
)
Reference
Actual Position
Fig. 2. Desired and Actual Piezoelectric displacement.
0 0.2 0.4 0.6 0.8 1
400
200
0
200
400
Time (Sec)
V
e
l
o
c
i
t
y
(
M
i
c
r
o
m
e
t
e
r
/
S
e
c
)
Fig. 3. Piezoelectric velocity.
A Robust Motion Tracking Control
of PiezoPositioning Mechanism with Hysteresis Estimation
107
As depicted from the results, especially Fig.4, the positioning mechanism has a good
accuracy and the error between displacement and reference signal at least 4 orders is smaller
than the actual signal which means that the error is about %0.01.
0 0.2 0.4 0.6 0.8 1
5
0
5
Time (Sec)
D
i
s
p
l
a
c
e
m
e
n
t
E
r
r
o
r
(
N
a
n
o
m
e
t
e
r
)
Fig. 4. Error between actual & desired displacement.
For the purpose of study the behavior of the controller in the presence of parameter
uncertainties, by considering the 0.1 M M A = , 0.1 D D A = and 0.4
L
F A = and using
(0) 15( ), (0) 0, (0) 0 x m x z = = = as the initial values, we have performed another test. The
result of this test is compared with the result of the method presented in [19]. In the
simulation we consider the control gain o is 2000.
As it can be seen from Fig.5 and Fig.7, the presented method has an acceptable response
and the positioning error is in about %0.25. Also, the hysteresis and disturbance voltage
and its estimation are shown in Fig.9. As it can be seen from this result, the hysteresis
identifier can estimate the hysteresis voltage precisely. For the purpose of comparison the
accuracy of the proposed method with the recently proposed method in [19], we simulate
the piezopositioning mechanism with the method of [19]. The displacement error of [19]
is shown in Fig.10. Noted that in the simulation of [19] we use these control gains:
5
10 , 9000, 50
p v s
k k k = = = and 10 o = . Comparison of Fig.7 and Fig.10 depicts that the
error in the presented method is smaller than the method of [19]. Also, it is clear that in
addition of smaller number of control gains, the gain of the controller in the presented
method in comparison with [19] is very smaller. Experimental implementation of the high
gain needs more complexity and also it may be generate noise. Although method of [19] is
Applications of Nonlinear Control
108
0 0.1 0.2 0.3 0.4 0.5
0
5
10
15
20
25
30
Time (Sec)
D
i
s
p
l
a
c
e
m
e
n
t
(
M
i
c
r
o
m
e
t
e
r
)
Reference
Actual Position
Fig. 5. Displacement of piezoelectric.
0 0.1 0.2 0.3 0.4 0.5
800
400
0
400
800
Time (Sec)
V
e
l
o
c
i
t
y
(
M
i
c
r
o
m
e
t
e
r
/
S
e
c
)
Fig. 6. Velocity of piezoelectric.
A Robust Motion Tracking Control
of PiezoPositioning Mechanism with Hysteresis Estimation
109
0 0.1 0.2 0.3 0.4 0.5
.15
.1
.05
0
.05
.1
.15
Time (Sec)
D
i
s
p
l
a
c
e
m
e
n
t
E
r
r
o
r
(
M
i
c
r
o
m
e
t
e
r
)
Fig. 7. Error between actual & desired displacement.
0 0.1 0.2 0.3 0.4 0.5
1
0.5
0
0.5
1
Time (Sec)
C
o
n
t
r
o
l
I
n
p
u
t
(
V
)
Fig. 8. Control input.
Applications of Nonlinear Control
110
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
1
0.5
0
0.5
1
H
y
s
t
e
r
e
s
i
s
&
D
i
s
t
u
r
b
a
n
c
e
V
o
l
t
a
g
e
(
V
)
Time (Sec)
Error
Fig. 9. Estimated & actual values of hysteresis and disturbance voltages and error between
them.
0 0.1 0.2 0.3 0.4 0.5
.5
0
.5
Time (Sec)
D
i
s
p
l
a
c
e
m
e
n
t
E
r
r
o
r
(
M
i
c
r
o
m
e
t
e
r
)
Fig. 10. Displacement Error of the method presented in [19].
A Robust Motion Tracking Control
of PiezoPositioning Mechanism with Hysteresis Estimation
111
robust against disturbances and uncertainties and in comparison with other methods has
higher precision, it needs high value gains to perform this task. While the proposed
method can perform these tasks with lower cost and also it has the capability of hysteresis
estimation.
For the comparison between the presented method and the method of [19] another
simulation test is carried out. In this case 0.2 M M A = , 0.2 D D A = and 0.4
L
F A = . The result is
shown in Fig.11.
As it can be depicted the presented method has the better response and using the proposed
controller the error is decreased. Note that in this case the reference is similar to the
Fig.5.
The last simulation is devoted to the reference signal of Fig.2 by considering 0.2 M M A = ,
0.2 D D A = and 0.4
L
F A = . Also in this case the error of the proposed controller is less than
the method of [19].
0 0.2 0.4 0.6 0.8 1
0.2
0.1
0
0.1
0.2
0.3
Time (Sec)
E
r
r
o
r
(
M
i
c
r
o
m
e
t
e
r
)
Error of [19]
Error of
Proposed
Method
Fig. 11. Comparison between the displacement error of the presented controller and the
controller proposed in [19].
Applications of Nonlinear Control
112
0.2 0.6 1 1.4 1.8
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
Time (Sec)
E
r
r
o
r
(
M
i
c
r
o
m
e
t
e
r
)
Error of
Proposed
Method
Error of [19]
Fig. 12. Comparison between the displacement error of the presented controller and the
controller proposed in [19].
7. Conclusion
A robust tracking motion controller is designed to control a positioning of piezoelectric
actuator system with hysteresis phenomenon. The Lugre hysteresis model is used to model
the nonlinearities in the system under study. Using the presented controller, we can
estimate the nonlinear hysteresis and disturbances imposed to the piezoelectric actuator and
compensate their effects. The performance and efficiency of the designed controller for a
positioning system is compared with recently proposed robust method. The results obtained
depict the validity and performance of the presented approach.
8. References
[1] Spanner, K. and S. Vorndran, 2003. Advances in piezonanopositioning technology.
Proc. IEEE/ASME int. conf. advanced intelligent mechatronics, Kobe,
Japan.
[2] Yi B.J., G.B. Chung, H.Y. Na, W.K. Kim and I.H. Suh, 2003. Design and experiment of a
3DOF parallel micromechanism utilizing flexurehinges. IEEE Trans Robotics
Automation, 19(4): 60412.
A Robust Motion Tracking Control
of PiezoPositioning Mechanism with Hysteresis Estimation
113
[3] EI Rifai, O.M., 2002. Modeling and control of undesirable dynamics in atomic force
microscopes. PhD dissertation, MIT, February.
[4] Hung, X.C. and D.Y. Lin, 2003. Tracking control of a piezoelectric actuator based on
experiment hysteretic modeling. Proc. Of the IEEE/ASME Int. Conf. Advanced
Intelligent Mechatronics, Kobe, Japan.
[5] Ge, P. and M. Jouaneh, 1997. Generalized priesach model for hysteresis nonlinearity of
piezoceramic actuators. Precision Eng., 20: 99111.
[6] Richter, H., E.A. Misawa, D.A. Lucca and H. Lu, 2001. Modeling nonlinear behavior in a
piezoelectric actuator. Prec. Eng. J., 25: 128137.
[7] Krasnosellskii, M.A. and A.V. Pokrovskii, 1989. Systems with hysteresis, Springer
Verlag, Berlin.
[8] Stepanenko, Y. and C.Y. Su, 1998. Intelligent control of piezoelectric actuators. Proc.
IEEE Conf. Decision & Control, 42344239.
[9] Wen, Y.K. 1976. Method for random vibration of hysteresis system. J. Eng. Mechanics
Division, 102: 249263.
[10] Zhou, J., C. Wen and C. Zhang, 2007. Adaptive backstepping control of piezo
positioning mechanism with hysteresis. Trans. CSME, 31(1): 97110.
[11] Yavari, F., M.J. Mahjoob, and C. Lucas, 2007. Fuzzy control of piezoelectric actuators
with hysteresis for nanopositioning. 13th IEEE/IFAC Int. Conf. on Methods and
Models in Automation and Robotics, 685690, August.
[12] Xu, J.H. 1993. Neural network control of a piezo tool positioner. Canadian Conf. Elect.
& Comp. Eng., 1: 333336.
[13] Jung, S.B. and S.W. Kim, 1994. Improvement of scanning accuracy of pzt piezoelectric
actuator by feedforward modelreference control. Precision Engineering, 16: 4955.
[14] Tao, G. and P.V. Koktokovic, 1995. Adaptive control of plants with unknown hysteresis.
IEEE Transaction on Auotomatic Control, 40: 200212.
[15] Huwang, L. and C. Jans, 2003. A reinforcement discrete neuroadaptive control of
unknown piezoelectric actuator systems with dominant hysteresis. IEEE Trans.
Neural Networks, 14: 6678.
[16] Lin, F.J., H.J. Shieh, and P.K. Huang, 2006. Adaptive wavelet neural network control
with hystereis estimation for piezopositioning mechanism. IEEE Trans. Neural
Networks, 17: 432444.
[17] Jan, C. and C.L. Hwang, 2004. A nonlinear observerbased slidingmode control for
piezoelectric actuator systems: theory and experiments. J Chinese Inst. Eng.,27(1):
922.
[18] Shieh H.J., F.J. Lin, P.K. Huang, and L.T. Teng, 2004. Adaptive tracking control solely
using displacement feedback for a piezopositioning mechanism. IEE Proc Control
Theory Appl,151(5): 65360.
[19] Liaw, H.C., B. Shirinzadeh, and J. Smith, 2008. Robust motion traking control of piezo
driven flexurebased four bar mehanism for micro/nano manipulation.
Mechatronics, 18: 111120.
[20] Yannier, S. and A. Sabanovic, 2007. Continuous Time Controller Based on SMC and
Disturbance Observer for Piezoelectric Actuators. Int. Rev. Elect. Eng (I.R.E.E.), 3 (6).
Applications of Nonlinear Control
114
[21] Spooner, J.T., M. Maggiore, R. Ordonez, and K.M. Passino, 2002. Stable Adaptive
Control and Estimation for Nonlinear Systems. John Wiley and Sons, Inc, NY.
0
Nonlinear ObserverBased
Control Allocation
Fang Liao
1
, Jian Liang Wang
2
and KaiYew Lum
1
1
National University of Singapore
2
Nanyang Technological University
Singapore
1. Introduction
Control allocation is the process of mapping virtual control inputs (such as torque and force)
into actual actuator deections in the design of control systems (Benosman et al., 2009;
Bodson, 2002; Bufngton et al., 1998; Liao et al., 2007; 2010). Essentially, it is considered as
a constrained optimization problemas one usually wants to fully utilize all actuators in order
to minimize power consumption, drag and other costs related to the use of control, subject to
constraints such as actuator position and rate limits. In the design of control allocation, full
state information is required. However, in practice, states may not be measurable. Hence,
estimation of these unmeasurable states becomes inevitable.
The unmeasurable states are generally estimated based on available measurements and
the knowledge of the physical system. For linear systems, the property of observability
guarantees the existence of an observer. Luenberger or Kalman observers are known to give
a systematic solution (Luenberger, 1964). In the case of nonlinear systems, observability in
general depends on the input of the system. In other words, observability of a nonlinear
system does not exclude the existence of inputs for which two distinct initial states generate
identical measured outputs. Hence, in general, observer gains can be expected to depend on
the applied input (Nijmeijer & Fossen, 1999). This makes the design of a nonlinear observer
for a general nonlinear system a challenging problem. Although various results have been
proposed over the past decades (AhmedAli & LamnabhiLagarrigue, 1999; Alamir, 1999;
Besancon, 2007; Besancon & Ticlea, 2007; Bestle & Zeitz, 1983; Bornard & Hammouri, 1991;
Gauthier & Kupka, 1994; Krener & Isidori, 1983; Krener & Respondek, 1985; Michalska &
Mayne, 1995; Nijmeijer & Fossen, 1999; Teel & Praly, 1994; Tsinias, 1989; 1990; Zimmer, 1994),
none of them can claim to provide a general solution with the same convergence properties as
in the linear case.
Over the past decades, a variety of methods have been developed for constructing nonlinear
observers for nonlinear systems (AhmedAli & LamnabhiLagarrigue, 1999; Alamir, 1999;
Besancon, 2007; Besancon & Ticlea, 2007; Bestle & Zeitz, 1983; Bornard & Hammouri, 1991;
Gauthier & Kupka, 1994; Krener & Isidori, 1983; Krener & Respondek, 1985; Michalska &
Mayne, 1995; Nijmeijer & Fossen, 1999; Teel & Praly, 1994; Tsinias, 1989; 1990; Zimmer,
1994). They may be classied into optimizationbased methods (Alamir, 1999; Michalska
7
2 WillbesetbyINTECH
& Mayne, 1995; Zimmer, 1994) and feedbackbased methods (Bestle & Zeitz, 1983; Bornard
& Hammouri, 1991; Gauthier & Kupka, 1994; Krener & Isidori, 1983; Krener & Respondek,
1985; Teel & Praly, 1994; Tsinias, 1989; 1990). Optimizationbased methods obtain an estimate
x(t) of the state x(t) by searching for the best estimate x(0) of x(0) (which can explain
the evolution y() over [0, t]) and integrating the deterministic nonlinear system from x(0)
and under u(). These methods take advantage of their systematic formulation, but suffer
from usual drawbacks of nonlinear optimization (like computation burden, local minima,
and so on). Feedbackbased methods can correct online the estimation x(t) from the
error between the measurement output and the estimated output. These methods include
linearization methods (Bestle & Zeitz, 1983; Krener & Isidori, 1983; Krener & Respondek,
1985), Lyapunovbased approaches (Tsinias, 1989; 1990), sliding mode observer approaches
(AhmedAli & LamnabhiLagarrigue, 1999) and high gain observer approaches (Bornard &
Hammouri, 1991; Gauthier & Kupka, 1994; Teel & Praly, 1994), and so on. Among them,
linearization methods (Krener &Isidori, 1983) transformnonlinear systems into linear systems
by change of state variables and output injection. It is applicable to a special class of nonlinear
systems. Sliding mode observer approaches (AhmedAli & LamnabhiLagarrigue, 1999) is
to force the estimation error to join a stabilizing variety. The difculty is to nd a variety
attainable and having this property. High gain observer approaches (Besancon, 2007) use the
uniform observability and weight a gain based on the linear part so as to make the linear
dynamics of the observer error to dominate the nonlinear one. Due to the requirement of the
uniform observability, these approaches can only be applied to a class of nonlinear systems
with special structure. Interestingly, Lyapunovbased approaches (Tsinias, 1989; 1990) provide
a general sufcient Lyapunov condition for the observer design of a general class of nonlinear
systems and the proposed observer is a direct extension of Luenberger observer in linear case.
In this chapter, we extend the control allocation approach developed in (Benosman et al., 2009;
Liao et al., 2007; 2010) from state feedback to output feedback and adopt the Lyapunovtype
observer for a general class of nonlinear systems in (Tsinias, 1989; 1990) to estimate the
unmeasured states. Sufcient Lyapunovlike conditions in the form of the dynamic update
law are proposed for the control allocation design via output feedback. The proposed
approach ensures that the estimation error and its rate converge exponentially to zero as
t + and the closedloop system exponentially converges to the stable reference model
as t +. The advantage of the proposed approach is that it is applicable to a wide class
of nonlinear systems with unmeasurable states, and it is computational efciency as it is not
necessary to optimize the control allocation problem exactly at each time instant.
This chapter is organized as follows. In Section 2, the observerbased control allocation
problem is formulated where the control allocation design is based on the estimated states
which exponentially converge to the true states as t +. In Section 3, the main result of
the observerbased control allocation design is presented in the form of dynamic update law.
An illustrative example is given in Section 4, followed by some conclusions in Section 5.
Throughout this chapter, given a real map f (v, w), (v, w) R
n
R
m
, D
v
f (v
0
, w
0
) denotes
its derivative with respect to v at the point (v
0
, w
0
). For given real map h(v) with v R
n
,
Dh(v
0
) denotes its derivative with respect to v at the point v
0
. In addition,   represent the
induced 2norm.
116 Applications of Nonlinear Control
Nonlinear ObserverBased
Control Allocation 3
2. Problem formulation
Consider the following nonlinear system:
_
x = f (x, u)
y = h(x)
(1)
where x A R
n
is the state vector with A a open subset of R
n
, y R
l
is the measurement
output vector, and u R
m
is the control input vector satisfying the constraints
u
=
_
u = [u
1
u
2
u
m
]
T
u
i
u
i
u
i
, i = 1, 2, , m
_
(2)
with u = [u
1
u
2
u
m
]
T
and u = [ u
1
u
2
u
m
]
T
being vectors of lower and upper
control limits, respectively.
We assume that the system (1) satises the following assumption:
Assumption 1. The function f (x, u) is smooth and the output function h(x) is continuously
differentiable.
Since control allocation need full state information, the state estimation for the system (1) is
required.
Consider a dynamic observer of the following form
T
( x, r, u)H
2
( x, r, u) (7)
where H
1
R
mm
and H
2
R
nn
are positive denite weighting matrices, and
( x, r, u)
= f ( x, u) A
d
x B
d
r (8)
is the matching error between the actual dynamics and desired dynamics. Since power
consumption minimization introduced by the term
1
2
u
T
H
1
u is a secondary objective, we
choose H
1
 H
2
.
Now the control allocation problem is formulated in terms of solving the following nonlinear
static minimization problem:
min
u
J( x, r, u) subject to
u and x converges to x exponentially
(9)
Dene
(u) = [S(u
1
) S(u
2
) S(u
m
)] (10)
with
S(u
i
)=min((u
i
u
i
)
3
, ( u
i
u
i
)
3
, 0), i = 1, 2, , m (11)
Then the constraint condition u is equivalent to
(u) = 0 (12)
Introduce the Lagrangian
L( x, r, u, )=J( x, r, u)+(u) (13)
where R
m
is a Lagrange multiplier. And assume that
Assumption 3. There exists a constant
1
> 0 such that
2
L
u
2
1
I
m
.
The following lemma is immediate ((Wismer & Chattergy, 1978), p. 42).
Lemma 1. If Assumptions 1 and 3 hold, the Lagrangian (13) achieves a local minimum if and only if
L
= 0 and
L
u
= 0.
118 Applications of Nonlinear Control
Nonlinear ObserverBased
Control Allocation 5
Proof. Necessity: The necessary condition is obvious. Sufciency: Since
L
= 0, we have
(u) = 0. In this case, the Lagrangian (13) is independent of the Lagrange multiplier , which
achieves a local minimum if
L
u
= 0 and
2
L
u
2
> 0. As
2
L
u
2
> 0 is guaranteed by Assumption
3, thus,
L
= 0 and
L
u
= 0 implies the local minimum. The proof is completed.
Remark 1. It should be noted that Assumption 3 is satised if all control inputs are within their
limits (i.e.,
L
=
T
(u) = 0) and the nonlinear system (1) is afne in control (i.e., f (x, u) =
f
1
(x) + g(x)u). It is because, in this case,
2
L
u
2
= H
1
+ g
T
(x)H
2
g(x) is positive denite matrix for
H
1
> 0 and H
2
> 0. Furthermore, since the Lagrangian (13) is convex in this case, Lemma 1 holds
for a global minimum.
To solve the control allocation problem (9) with the state estimate x from the observer (3), we
consider the following control Lyapunovlike function
V( x, e, r, u, ) = V
m
( x, r, u, ) +
1
2
e
T
Pe (14)
where P > 0 is a known positivedenite matrix and
V
m
( x, r, u, )=
1
2
_
_
L
u
_
T
L
u
+
_
L
_
T
L
_
(15)
Here the function V
m
is designed to attract (u, ) so as to minimize the Lagrangian (13). The
term
1
2
e
T
Pe forms a standard Lyapunovlike function for observer estimation error e which
is required to exponentially converge to zero as t +.
Following the observer design in (Tsinias, 1989), we dene a neighborhood Q of zero with
Q A, a neighborhood W of A with x e : x A, e Q W, and a closed ball S of
radius r > 0, centered at zero, such that S Q. Then dene the boundary of S as S. Figure 1
illustrates the geometrical relationship of these dened sets.
Let 1 denote the set of the continuously differentiable output mappings h(x) : A R
l
such
that for every m
0
Q and x W,
R( x, m
0
) 0 (16)
and
kerR( x, m
0
) kerDh( x) (17)
where
R( x, m
0
)
= [Dh( x)]
T
Dh( x + m
0
) + [Dh( x + m
0
)]
T
Dh( x) (18)
Remark 2. Obviously, every linear map y = Hx belongs to 1. Furthermore, 1 contains a wide
family of nonlinear mappings.
119 Nonlinear ObserverBased Control Allocation
6 WillbesetbyINTECH
r
S
W
Q
s e
X
0
Fig. 1. Geometrical representation of sets
We assume that
Assumption 4. h(x) in the system (1) belongs to the set 1, namely, h(x) 1.
Further, we dene
N
=
_
e R
n
x
e
T
PD
x
f ( x + m
1
, u)e k
0
e
2
_
(19)
and assume that
Assumption 5. There exist a positive denite matrix P R
n
x
n
x
and a positive constant k
0
such
that kerDh( x) N holds for any ( x, m
1
, u) WQ.
Remark 3. Assumption 5 ensures that the estimation error system (5) is stable in the case of h(x) =
h( x) and x ,= x. In particular, for linear systems, the condition in Assumption 5 is equivalent to
detectability.
3. Main results
Denote
_
_
=
2
L
u
2
2
L
u
2
L
u
0
L
u
L
(20)
and dene
M
=
_
R
n
x
= re
1
e, e N S
_
(21)
Let
1
( x, u) = max
_
r
2
(PD
x
f ( x + m
1
, u) + k
0
), m
1
S, ( x, u) W
_
(22)
2
( x) = min
_
1
2
T
R( x, m
0
), m
0
S, S M, x W
_
(23)
120 Applications of Nonlinear Control
Nonlinear ObserverBased
Control Allocation 7
Theorem 1. Consider the system (1) with x A and u . Suppose that Assumptions
15 are satised. For a given asymptotically stable matrix A
d
and a matrix B
d
, given symmetric
positivedenite matrices
1
and
2
, and a given positive constants , for e(0) near zero,
_
L
,
L
u
, e
_
exponentially converges to zero as t +, and the dynamics of the nonlinear system(1) exponentially
converges to that of the stable system (6) if the following dynamic update law
_
u =
1
+
1
=
2
+
2
(24)
and the observer system
x = f ( x, u) ( x, u) [y h( x)] (25)
are adopted. Here , R
m
are as in (20), and
1
,
2
R
m
satisfy
1
+
T
2
+ + V
m
= 0 (26)
with V
m
as in (15) and
=
_
L
u
_
T
2
L
ru
r +
_
L
u
_
T
2
L
xu
x (27)
and the mapping
( x, u) = ( x, u)P
1
[Dh( x)]
T
(28)
where
( x, u)
1
( x, u)
2
( x)
> 0 (29)
with
1
( x, u) > 0 and
2
( x) > 0 dened as in (22) and (23).
Proof. From the Lyapunovlike function (14), we obtain its time derivative as
V =
_
_
L
u
_
T
2
L
u
2
+
_
L
_
T
2
L
u
_
u +
_
L
u
_
T
2
L
u
+
_
L
u
_
T
2
L
ru
r +
_
L
u
_
T
2
L
xu
x + e
T
P e (30)
Substituting e in (5), and as in (20) and as in (27) into (30), we have
V =
T
u +
T
+ + e
T
P f (x, u) f ( x, u) +( x, u)[y h( x)] (31)
Consider e S. Since S is convex, according to Mean Value Theorem, there exists m
0
, m
1
S
satisfying
f (x, u) f ( x, u) = D
x
f ( x + m
1
, u)e (32)
y h( x) = Dh( x + m
0
)e (33)
121 Nonlinear ObserverBased Control Allocation
8 WillbesetbyINTECH
Then substituting (24), (26), (32) and (33) into (31), we obtain
V =
T
1
T
2
V
m
+ e
T
P [D
x
f ( x + m
1
, u) +( x, u)Dh( x + m
0
)] e (34)
After substituting ( x, u) as in (28) and R( x, m
0
) as in (18), (34) can be rewritten as
V =
T
1
T
2
V
m
+ e
T
PD
x
f ( x + m
1
, u)e
( x, u)
2
e
T
R( x, m
0
)e (35)
Since the matrices
1
> 0 and
2
> 0, we have
V V
m
+ e
T
PD
x
f ( x + m
1
, u)e
( x, u)
2
e
T
R( x, m
0
)e (36)
For e = 0 where x is determined by the observer accurately, we have
V V
m
= V (37)
Since > 0, V exponentially converges to zero as t +. Hence,
_
L
,
L
u
_
exponentially
converges to zero.
For any nonzero e S, let = re
1
e. Obviously, S. Then we have
V V
m
+
1
r
2
e
2
T
PD
x
f ( x + m
1
, u)
( x, u)
2r
2
e
2
T
R( x, m
0
) (38)
In the following, we shall show that V converges exponentially to zero for all m
0
, m
1
S,
x W, u , e S, e ,= 0 and S.
First let us consider nonzero e N S. From = re
1
e, we have M. Since m
0
,
m
1
S Q, x W and u , according to Assumptions 15, it follows that
T
R( x, m
0
) = 0 (39)
and
T
PD
x
f ( x + m
1
, u) k
0

2
(40)
with the constant k
0
> 0. From (38), we have
V V
m
k
0
e
2
V (41)
with the constant > 0. Hence,
_
L
,
L
u
, e
_
exponentially converges to zero as t +.
Then we consider nonzero e S N S, namely, S M. From (38), taking into account
(22)(23), we obtain
V V
m
+
1
r
2
e
2
_
1
( x, u) k
0
r
2
( x, u)
2
( x)
_
(42)
Since ( x, u) satisfy the condition (29), we obtain (41) again. Hence, in this case,
_
L
,
L
u
, e
_
also exponentially converges to zero as t +.
122 Applications of Nonlinear Control
Nonlinear ObserverBased
Control Allocation 9
Since
_
L
,
L
u
, e
_
exponentially converges to zero as t +, the closedloop system
exponentially converges to
_
x = A
d
x + B
d
r
e =
_
D
x
f ( x + m
1
, u) ( x, u)P
1
[Dh( x)]
T
Dh( x + m
0
)
_
e
(43)
Since A
d
is a asymptotically stable matrix, we know that x W is bounded. According to
Assumptions 1 and 4, D
x
f ( x + m
1
, u), Dh( x) and Dh( x + m
0
) are all bounded for m
0
, m
1
S
and u . From k
0
> 0, we have 0 <
1
( x, u) < +. According to Assumption 4, we have
kerR( x, m
0
) kerDh( x) which ensures that 0 <
T
R( x, m
0
) < + for every S M,
m
0
S and x W. Thus, we have 0 <
2
( x) < +. As a result, 0 < ( x, u) < +. From
(43), we know that e exponentially converges to zero as e exponentially converges to zero.
Moreover, we have
x e = A
d
x A
d
e + B
d
r (44)
Since e and e exponentially converges to zero, we have the system(1) exponentially converges
to x = A
d
x + B
d
r. This completes the proof.
Consider now the issue of solving (26) with respect to
1
and
2
. One method to achieve
a welldened unique solution to the underdetermined algebraic equation is to solve a
leastsquare problem subject to (26). This leads to the Lagrangian
l(
1
,
2
, ) =
1
2
(
T
1
1
+
T
2
2
) + (
T
1
+
T
2
+ + V
m
) (45)
where R is a Lagrange multiplier. The rst order optimality conditions
l
1
= 0,
l
2
= 0,
l
= 0 (46)
leads to the following system of linear equations
I
m
0
0 I
m
T
0
0
0
V
m
(47)
Remark 4. It is noted that Equation (47) always has a unique solution for
1
and
2
if any one of
and is nonzero.
4. Example
Consider the pendulum system
_
x
1
x
2
_
=
_
x
2
sin x
1
+ u
1
cos x
1
+ u
2
sin x
1
_
(48)
y = x
1
+ x
2
(49)
123 Nonlinear ObserverBased Control Allocation
10 WillbesetbyINTECH
with x = [x
1
x
2
]
T
R
2
, u = [u
1
u
2
]
T
and
=
_
u = [u
1
u
2
]
T
1 u
1
1, 0.5 u
2
0.5
_
(50)
As the system is afne in control and its measurement output y is a linear map of its state x,
Assumptions 1, 3 and 4 are satised automatically.
Choose
P =
_
3 0
0 1
_
For e ,= 0 and e ker[1 1], we have e
1
= e
2
and
e
T
PD
x
f (x, u)e[
e
1
=e
2
= [e
1
e
2
]
_
0 3
cos x
1
u
1
sin x
1
+ u
2
cos x
1
0
_ _
e
1
e
2
_
[
e
1
=e
2
= (cos x
1
u
1
sin x
1
+ u
2
cos x
1
+3)e
1
e
2
[
e
1
=e
2
[1.5 cos(arctan
2
3
) sin(arctan
2
3
) +3]e
1
e
2
[
e
1
=e
2
= (1.8028 +3)e
1
e
2
[
e
1
=e
2
= 0.5986e
2
[
e
1
=e
2
< k
0
e
2
[
e
1
=e
2
with 0 < k
0
< 0.5986. Hence, Assumption 5 is satised. Let S be the ball of radius r = 1,
centered at zero and S is the boundary of S. Dene M S and
M =
_
= [
1
2
]
T
R
2
:  = 1, 3
1
2
+1.8028[
1
2
[ < k
0
_
Obviously,
S M =
_
= [
1
2
]
T
R
2
:  = 1, 3
1
2
+1.8028[
1
2
[ k
0
_
As
1
( x, u) = 3 1.8028 + k
0
and
2
( x) = min
_
(
1
+
2
)
2
, S M
_
= 1
2k
0
3 1.8028
choosing k
0
= 0.5, we have
1
( x, u)
2
( x)
= 35.8699. Let ( x, u) = 36 > 35.8699 and we have
( x, u) = [12 36]
T
.
Now the nonlinear observer becomes
_
x
1
x
2
_
=
_
x
2
sin x
1
+ u
1
cos x
1
+ u
2
sin x
1
_
+
_
12
36
_
(y x
1
x
2
)
Choose the reference model (6) where
A
d
=
_
0 1
25 10
_
, B
d
=
_
0
25
_
124 Applications of Nonlinear Control
Nonlinear ObserverBased
Control Allocation 11
and the reference is given by
r =
r
f
_
6
_
t
t
1
_
5
15
_
t
t
1
_
4
+10
_
t
t
1
_
3
_
, 0 t < t
1
r
f
, t
1
t < t
2
r
f
6
_
t t
2
t
f
t
2
_
5
15
_
t t
2
t
f
t
2
_
4
+10
_
t t
2
t
f
t
2
_
3
+ r
f
, t
2
t < t
f
0, t t
f
with t
1
= 10s, t
2
= 20s, t
f
= 30s and r
f
= 0.5. Obviously, Assumption 2 is satised.
Set H
1
= 0, H
2
= 10
4
I
2
, = 1,
1
=
2
= 2I
2
, and x
1
(0) = 0.3 and x
2
(0) = 0.5. Using the
proposed approach, we have the simulation result of the pendulum system(48)(50) shown in
Figures 25 where the control u
2
is stuck at 0.5 from t = 12s onward.
From Figure 2, it is observed that the estimated states x
1
and x
2
converge to the actual
states x
1
and x
2
and match the desired states x
1d
and x
2d
well, respectively, even when
u
2
is stuck at 0.5. This observation is further veried by Figure 3 where both the state
estimation errors e
1
(= x
1
x
1
) and e
2
(= x
2
x
2
) of the nonlinear observer as in (4) and
the matching errors
1
(= 0) and
2
(= sin x
1
+ u
1
cos x
1
+ u
2
sin x
1
+25 x
1
+10 x
2
25r) as
in (8) exponentially converge to zero. Moreover, Figure 4 shows that the control u
1
roughly
satises the control constraint u
1
[1, 1] while the control u
2
strictly satises the control
constraint u
2
[0.5, 0.5]. This is because, in this example, the Lagrange multiplier
1
is rst
activated by the control u
1
< 1 at t = 0 (see Figure 5 where
1
is no longer zero from t = 0),
0 5 10 15 20 25 30
0
0.2
0.4
time(s)
s
t
a
t
e
v
a
r
i
a
b
l
e
1
0 5 10 15 20 25 30
0.5
0
0.5
time(s)
s
t
a
t
e
v
a
r
i
a
b
l
e
2
hx
1
x
1
x
1d
hx
2
x
2
x
2d
Fig. 2. Responses of the desired, estimated and actual states
125 Nonlinear ObserverBased Control Allocation
12 WillbesetbyINTECH
0 5 10 15 20 25 30
0.2
0
0.2
0.4
0.6
time(s)
e
s
t
i
m
a
t
i
o
n
e
r
r
o
r
0 5 10 15 20 25 30
5
0
5
10
time(s)
m
a
t
c
h
i
n
g
e
r
r
o
r
e
1
e
2
1
2
Fig. 3. Responses of estimation error and matching error
0 5 10 15 20 25 30
1
0.5
0
0.5
1
time(s)
u
1
0 5 10 15 20 25 30
0.4
0.2
0
0.2
0.4
0.6
time(s)
u
2
Fig. 4. Responses of control u
and then the proposed dynamic update law forces the control u
1
to satisfy the constraint
u
1
[1, 1]. It is also noted from Figure 5 that the Lagrange multiplier
2
is not activated in
this example as the control u
2
is never beyond the range [0.5, 0.5]. In addition, the output y
and the Lyapunovlike function V
m
are shown in Figure 6. From Figure 6, it is observed that
the Lyapunovlike function V
m
exponentially converges to zero.
126 Applications of Nonlinear Control
Nonlinear ObserverBased
Control Allocation 13
0 5 10 15 20 25 30
0.5
0.4
0.3
0.2
0.1
0
0.1
time(s)
1
0 5 10 15 20 25 30
0.1
0.05
0
0.05
0.1
time(s)
2
Fig. 5. Responses of Lagrangian multiplier
0 5 10 15 20 25 30
0.5
0
0.5
1
time(s)
y
0 5 10 15 20 25 30
0
1
2
3
x 10
8
time(s)
V
m
Fig. 6. Responses of output y and Lyapunovlike function V
m
5. Conclusions
Sufcient Lyapunovlike conditions have been proposed for the control allocation design via
output feedback. The proposed approach is applicable to a wide class of nonlinear systems.
As the initial estimation error e(0) need be near zero and the predened dynamics of the
127 Nonlinear ObserverBased Control Allocation
14 WillbesetbyINTECH
closedloop is described by a linear stable reference model, the proposed approach will
present a local nature.
6. References
AhmedAli, T. & LamnabhiLagarrigue, F. (1999). Sliding observercontroller design
for uncertain triangle nonlinear systems, IEEE Transactions on Automatic Control
44(6): 12441249.
Alamir, M. (1999). Optimizationbased nonlinear observer revisited, International Journal of
Control 72(13): 12041217.
Benosman, M., Liao, F., Lum, K. Y. & Wang, J. L. (2009). Nonlinear control allocation
for nonminimum phase systems, IEEE Transactions on Control Systems Technology
17(2): 394404.
Besancon, G. (ed.) (2007). Nonlinear Observers and Applications, Springer.
Besancon, G. & Ticlea, A. (2007). An immersionbased observer design for rankobservable
nonlinear systems, IEEE Transactions on Automatic Control 52(1): 8388.
Bestle, D. & Zeitz, M. (1983). Canonical form observer design for nonlinear timevariable
systems, International Journal of Control 38: 419431.
Bodson, M. (2002). Evaluation of optimization methods for control allocation, Journal of
Guidance, Control and Dynamics 25(4): 703711.
Bornard, G. & Hammouri, H. (1991). A high gain observer for a class of uniformly observable
systems, Proceedings of the 30th IEEE Conference on Decision and Control, pp. 14941496.
Bufngton, J. M., Enns, D. F. &Teel, A. R. (1998). Control allocation and zero dynamics, Journal
of Guidance, Control and Dynamics 21(3): 458464.
Gauthier, J. P. & Kupka, I. A. K. (1994). Observability and observers for nonlinear systems,
SIAM Journal on Control and Optimization 32: 975994.
Krener, A. J. & Isidori, A. (1983). Linearization by output injection and nonlinear observers,
Systems & Control Letters 3: 4752.
Krener, A. J. & Respondek, W. (1985). Nonlinear observers with linearizable error dynamics,
SIAM Journal on Control and Optimization 23: 197216.
Liao, F., Lum, K. Y., Wang, J. L. & Benosman, M. (2007). Constrained nonlinear nitetime
control allocation, Proc. of 2007 American Control Conference, New York City, NY.
Liao, F., Lum, K. Y., Wang, J. L. & Benosman, M. (2010). Adaptive Control Allocation
for Nonlinear Systems with Internal Dynamics, IET Control Theory & Applications
4(6): 909922.
Luenberger, D. G. (1964). Observing the state of a linear system, IEEE Transactions on Military
Electron 8: 7480.
Michalska, H. & Mayne, D. Q. (1995). Moving horizon observers and observerbased control,
IEEE Transactions on Aotomatic Control 40(6).
Nijmeijer, H. & Fossen, T. I. (eds) (1999). New Directions in Nonlinear Observer Design, Springer.
Teel, A. & Praly, L. (1994). Global stabilizability and observability imply semiglobal
stabilizability by output feedback, Systems & Control Letters 22: 313325.
Tsinias, J. (1989). Observer design for nonlinear systems, Systems & Control Letters 13: 135142.
Tsinias, J. (1990). Further results on the observer design problem, Systems & Control Letters
14: 411418.
Wismer, D. A. & Chattergy, R. (1978). Introduction To Nonlinear Optimization: A Problem Solving
Approach, Elsevier NorthHolland, Inc.
Zimmer, G. (1994). State observation by online minimization, International Journal of Control
60(4): 595606.
128 Applications of Nonlinear Control
8
Predictive Function Control of the
SingleLink Manipulator with Flexible Joint
Zhihuan Zhang and Chao Hu
Ningbo Institute of Technology, Zhejiang University
China
1. Introduction
Flexiblelink robotic manipulators have many advantages with respect to conventional rigid
robots. They are built by lighter, cheaper materials, which improve the payload to arm
weight ratio, thus resulting in an increase of the speed with lower energy consumption.
Moreover, due to the reduced inertia and compliant structure, these lightweight arms can be
operated more safely and are more applicable for the delicate assembly tasks and interaction
with fragile objects, including human beings.
The control for robot manipulators is to determine the time history of joint inputs to cause
the endeffector to execute a commanded motion. There are many control techniques and
methodologies that can be applied to the control of the manipulators. The specific control
method and its implementation ways can have a significant impact on the performance of
the manipulator and consequently on the range of its possible applications. In addition, the
mechanical design of the manipulator itself will influence the type of control scheme
needed. However, in order to improve the control performance, more sophisticated
approaches should be found.
The control for flexible joint system has attracted a considerable amount of attention during
the past few years. There are PD, inverse dynamics, and the force control approach for the
feedback control strategies of flexible joint manipulator. (1989, MARK W. SPONG), an
integral manifold approach to the feedback control of flexible joint robots (1987, MARK W.
SPONG, KHASHAYAR KHORASANI, and PETAR V. KOKOTOVIC), and the nonlinear
feedback control of flexible joint manipulators: a single link case study, (1990, K.
KHORASANI). The basic idea of feedback linearization is to construct a nonlinear control
law as a socalled inner loop control which, in the ideal case, exactly linearizes the nonlinear
system after a suitable state space change of coordinates. The designer can design a second
stage or outer loop control in the new coordinates to satisfy the traditional control design
specifications such as tracking, disturbance rejection, and so forth. Since the feedback
linearization of flexible joint manipulator is a fourth order integrator system, so we
proposed a three stage design method, the first is nonlinear feedback to get integrator
system, the second is pole placement to get expect performance, and the third is to use PFC
to reject disturbance and uncertainty, since they can not be exactly cancelled by nonlinear
feedback, coupling effects of the joint flexibility. More accurate description of robot
dynamics may include fast actuator dynamics and jointlink flexibility, and so on.
Applications of Nonlinear Control
130
2. Equations of motion
Consider the singlelink arm shown in Figure 1 consisting of a flexible joint.
Fig. 1. Singlelink robot with joint flexibility
The kinetic energy of the manipulator is a quadratic function of the vector q
,
1 1
( ) ( )
2 2
n
T
ij i j
i j
K q D q q d q q q = =
(1)
where the n n inertia matrix ( ) D q is symmetric and positive definite for each
n
q e9 .
The potential energy ( ) V V q = is independent of q . We have remarked that robotic
manipulator satisfies this condition.
The EulerLagrange equations for such a system can be derived as follows. Since
,
1
( ) ( )
2
n
ij i j
i j
L K V d q q q V q = =
(2)
we have
( )
kj j
k j
L
d q q
q
c
=
c
and
,
( ) ( ) ( )
kj
kj j kj j kj j i j
k i j j j i j
d
d L d
d q q d q q d q q q q
dt q dt q
c
c
= + = +
c c
Also
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
131
,
1
2
ij
i j
k k k i j
d
L V
q q
q q q
c
c c
=
c c c
Thus the EulerLagrange equations can be written as
,
1
( )
2
kj ij
kj j i j k
i k k j i j
d d
V
d q q q q
q q q
t
c c
c
+ =
`
c c c
)
, 1, , k n = (3)
By interchanging the order of summation and taking advantage of symmetry, we can show
that
, ,
1
2
kj kj
ki
i j i j
i i j i j i j
d d
d
q q q q
q q q
c c
c
= +
` `
c c c
)
)
Hence
, ,
1 1
2 2
kj ij kj ij
ki
i j i j
i k i j k i j i j
d d d d
d
q q q q
q q q q q
c c c c
c
= +
` `
c c c c c
)
)
The term
1
2
kj ij
ki
ijk
i j k
d d
d
c
q q q
c c
c
= +
`
c c c
)
(4)
are known as Christoffel symbols. Note that, for a fixed k, we have
ijk jik
c c = , which reduces
the effort involved in computing these symbols by a factor of about one half. Finally, if we
define
k
k
V
q
c
=
c
(5)
then the EulerLagrange equations can be written as
,
( ) ( ) ( )
kj j ijk i j k k
j i j
d q q c q q q q t + + =
, 1, , k n = (6)
In the above equation, there are three types of terms. The first involve the second derivative
of the generalized coordinates. The second are quadratic terms in the first derivatives of q,
where the coefficients may depend on q. These are further classified into two types. Terms
involving a product of the type
2
i
q are called centrifugal, while those involving a product of
the type
i j
q q where i j = are called Coriolis terms. The third type of terms are those
involving only q but not its derivatives. Clearly the latter arise from differentiating the
potential energy. It is common to write (6) in matrix form as
( ) ( , ) ( ) D q q C q q q g q t + + = (7)
Applications of Nonlinear Control
132
where the k, jth element of the matrix ( , ) C q q is defined as
1 1
1
( )
2
n n
kj ij
ki
kj ijk i i
i j k i i
d d
d
c c q q q
q q q
= =
c c
c
= = +
`
c c c
)
3. Feedback linearization design for inner loop
We first derive a model similar to (6) to represent the dynamics of a single link robot with joint
flexibility. For simplicity, ignoring damping of the equations of motion, system is given by
1 1 1 1 1 1 2
( ) ( , ) ( ) 0 D q q h q q q K q q + + = (8)
2 1 2
( ) Jq K q q u = (9)
In state space, which is now
4n
9 , we define state variables in block form
1 1 2 1
3 2 4 2
x q x q
x q x q
= =
= =
(10)
Then from (8)(9) we have
1 2
x x =
(11)
{ }
1
2 1 1 2 1 3
( ) ( , ) ( ) x D x h x x K x x
= +
(12)
3 4
x x =
(13)
1 1
4 1 3
( ) x J K x x J u
= +
(14)
This system is then of the form
( ) ( ) x f x G x u = +
(15)
For a singleinput nonlinear system, ( ) f x and ( ) g x are smooth vector fields on
n
9 ,
(0) 0 f = , and ue9, is said to be feedback linearizable if there exists a region U in
n
9 containing the origin, a diffeomorphism T:
n
U 9 , and nonlinear feedback
( ) ( ) u x x v o  = +
(16)
with ( ) 0 x  = on U, such that the transformed variables
( ) y T x =
(17)
satisfy the system of equations
y Ay bv = +
(18)
where
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
133
0 1 0 0
0 0 1 .
. . . . .
.
. . . . .
. . . 1
0 0 . . 0 0
A
(
(
(
(
=
(
(
(
(
(
0
0
.
.
.
1
b
(
(
(
(
=
(
(
(
(
(
In the singlelink case we see that the appropriate state variables with which to define the
system so that it could be linearized by nonlinear feedback on the link position, velocity,
acceleration, and jerk. Following the singleinput case, then, we can apply same action
on the multilink case and derive a feedback linearizing transformation blockwise as
follows,
1 1 1
( ) y T x x = = (19)
2 2 1 2
( ) y T x y x = = = (20)
{ }
1
3 3 2 2 1 1 2 1 3
( ) ( ) ( , ) ( ) y T x y x D x h x x K x x
= = = = + (21)
{ }
1 1
4 4 3 1 1 2 1 3 1 2
1
1 1
1 1 2 1 3 2 4 4 1 2 3 1 4
2
( ) [ ( ) ] ( , ) ( ) ( ) {
[ ( ) ( ( , ) ( ))] ( )} ( , , ) ( )
d h
y T x y D x h x x K x x D x x
dt x
h
D x h x x K x x K x x a x x x D x Kx
x
c
= = = +
c
c
+ + + = +
c
(22)
where for simplicity we define the function
4
a to be that in the definition of
4
y except the
last term, which is
1
4
D Kx
. Note that
4
x appears only in this last term so that
4
a depends
only on
1 2 3
, , x x x .
As in the singlelink case, the above mapping is a global diffeomorphism. Its inverse can be
found by
1 1
x y =
(23)
2 2
x y =
(24)
1
3 1 1 3 1 2
( ( ) ( , )) x y K D y y h y y
= + + (25)
1
4 1 4 4 1 2 3
( )( ( , , )) x K D y y a y y y
= (26)
The linearizing control law can now be found from the condition
4
y v =
(27)
where v is a new control input. Computing
4
y from (22) and suppressing function
arguments for brevity yields
Applications of Nonlinear Control
134
1 1 1 1 1 4 4 4
2 1 3 4 4 1 3
1 2 3
( ( )) [ ] ( ( ) )
( ) ( )
a a a d
v x D h K x x x D Kx D K J K x x J u
x x x dt
a x b x u
c c c
= + + + + +
c c c
= +
(28)
where
1 1 1 1 4 4 4
2 1 3 4 4 1 3
1 2 3
( ) : ( ( )) [ ] ( )
a a a d
a x x D h K x x x D Kx D KJ K x x
x x x dt
c c c
= + + + +
c c c
(29)
1 1
( ) ( ) b x D x KJ u
=
(30)
Solving the above expression for u yields
1
( ) ( ( )) u b x v a x
= (31)
: ( ) ( ) x x v o  = + (32)
where
1
( ) ( ) x JK D x 
= and
1
( ) ( ) ( ) x b x a x o
=
With the nonlinear change of coordinates (19)(22) and nonlinear feedback (32) the
transformed system has the linear block form
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
I
I
y y v
I
I
( (
( (
( (
= +
( (
( (
(33)
: Ay bv = +
where I n n = identity matrix, 0 n n = zero matrix,
4
1 2 3 4
( , , , )
T T T T T n
y y y y y = e9 , and
n
ve9 . The system (33) represents a set of n decoupled quadruple integrators.
4. Outer loop design based on predictive function control
4.1 why use predictive function control
The technique of feedback linearization is important due to it leads to a control design
methodology for nonlinear systems. In the context of control theory, however, one should
be highly suspicious of techniques that rely on exact mathematical cancellation of terms,
linear or nonlinear, from the equations defining the system.
In this section, we investigate the effect of parameter uncertainty, computational error,
model simplification, and etc. We show that the most important property of feedback
linearizable systems is not necessarily that the nonlinearities can be exactly cancelled by
nonlinear feedback, but rather that, once an appropriate coordinate system is found in
which the system can be linearized, the nonlinearities are in the range space of the input.
This property is highly significant and is exploited by the predictive function control
techniques to guarantee performance in the realistic case that the nonlinearities in the
system are not known exactly.
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
135
Consider a singleinput feedback linearizable system. After the appropriate coordinate
transformation, the system can be written in the ideal case as
1 2
1
( )[ ( )]
n
y y
y v x u x  o
=
= =
(34)
provided that u is given by (16) in order to cancel the nonlinear terms ( ) x o and ( ) x  .
In practice such exact cancellation is not achievable and it is more realistic to suppose that
the control law u in (16) is of the form
( ) ( ) u x x v o  = + (35)
where
( ) x o ,
( ) x o ,
( )[ ( ) ( ) ( )]
( , , , )
n
n
y y
y v x x x v x
v y y v
 o  o
q
=
= = +
= +
(36)
where the uncertainty q is given as
( )
( )
{ }
1
1 1
1
( )
( , , , ) 1 
n
y T X
y y v v q    o o
=
= + (37)
The system (36) can be written in matrix form as
{ ( , )} y Ay b v y v q = + + (38)
where A and b are given by (18). For multiinput case, similar to (33), and if
m
ve9 , and
q :
n m m
9 9 9 . Note that the system (38) is still nonlinear whenever 0 q = . The practical
implication of this is solved by the outer loop predictive function control (PFC).
The system (38) can be represented by the block diagram of Figure 2. The application of the
nonlinear inner loop control law results in a system which is approximately linear. A
common approach is to decompose the control input v in (38) into two parts, the first to
Applications of Nonlinear Control
136
stabilize the nominal linear system represented by (38) with 0 q = . In this case v can be
taken as a linear state feedback control law designed to stabilize the nominal system and/or
for tracking a desired trajectory. A second stage control v A is then designed for robustness,
that is, to guarantee the performance of the nominal design in the case that 0 q = . Thus the
form of the control law is
( ) ( ) u x x v o  = +
(39)
v Ky v = + A
(40)
( )
r
v PFC y A =
(41)
where Ky is a linear feedback designed to place the eigenvalues of A in a desired location,
v A represents an additional feedback loop to maintain the nominal performance despite the
presence of the nonlinear term q .
r
y is a reference input, which can be chosen as a signal for
tracking a desired trajectory.
Fig. 2. block diagram for PFC outer loop design
4.2 Predictive function control
All MPC strategies use the same basic approach i.e., prediction of the future plant outputs,
and calculation of the manipulated variable for an optimal control. Most MPC strategies are
based on the following principles:
Use of an internal model
Its formulation is not restricted to a particular form, and the internal model can be linear,
nonlinear, state space form, transfer function form, first principles, blackbox etc. In PFC,
nonlinear
system
of joint
flexibility
T(x)
( ) x o
( ) x 
PFC
linear model
base
function
x
y
y
u
v
e
y
K
q
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
137
only independent models where the model output is computed only with the present and
past inputs of the process models are used.
Specification of a reference trajectory
Usually an exponential.
Determination of the control law
The control law is derived from the minimization of the error between the predicted output
and the reference with the projection of the Manipulated Variable (MV) on a basis of functions.
Although based on these principles the PFC algorithm may be of several levels of
complexity depending on the order and form of the internal model, the order of the basis
function used to decompose the MV and the reference trajectory used.
4.3 First order PFC
Although it is unrealistic to represent industrial systems by a first order system, as most of
them are in a higher order, some well behaved ones may be estimated by a first order. The
estimation will not be perfect at each sample time, however, the robustness of the PFC will
help to maintain a decent control.
If the system can be modelled by a first order plus pure time delay system, then the
following steps in the development of the control law are taken.
Model formulation
In order to implement a basic first order PFC, a typical first order transfer function equation
(42) is used.
( ) ( )
1
M
M
M
K
y s u s
T S
=
+
(42)
Note that the time delay is not considered in the internal model formulation and in this case
M
K is equal to one. The discrete time formulation of the model zeroorder hold equivalent
is then obtained in (43).
( ) ( 1) (1 ) ( 1)
M M M
y k y k K u k o o = + (43)
where exp( )
s
M
T
T
o = . If the manipulated variable is structured as a step basis function:
( ) ( )
H
L M
y k H y k o + = (44)
( ) (1 ) ( )
H
F M
y k H K u k o + = (45)
Where,
L
y and
F
y are respectively, the free (autoregressive) and the forced response of
M
y .
Reference trajectory formulation
If
R
y is the expression of the reference trajectory, then at the coincidence point H:
Applications of Nonlinear Control
138
( ) ( ) ( ( ) ( ))
H
R P
C k H y k H C k y k + + = (46)
thus:
( ) ( ) ( ( ) ( ))
H
R P
y k H C k C k y k + = (47)
Predicted process output
The predicted process output is given by the model response, plus a term given the error
between the same model output and the process output:
( ) ( ) ( ( ) ( ))
P M P M
y k H y k H y k y k + = + + (48)
where ( ) ( ) ( ) ( ) (1 ) ( )
H H
M L F M M
y k H y k H y k H y k K u k o o + = + + + = + .
Computation of the control law
At the coincidence point H:
( ) ( )
R P
y k H y k H + = + (49)
Combining (44), (45), (47) and (48) yields
( ) ( ( ) ( )) ( ) ( ) ( )
H
P P M M
C k C k y k y k y k H y k = + (50)
Replacing ( )
M
y k H + by its equivalent in equations (44) and (45) we obtain:
( )(1 ) ( )(1 ) ( )(1 ) (1 ) ( )
H H H H
P M M
C k y k y k K u k o o + = (51)
Solving for u(k) the final result is the control law given in (52).
( ( ) ( ))(1 ) ( )
( )
(1 )
H
P M
H
M M
C k y k y k
u k
K K
o
= +
(52)
4.4 Case of a process with a pure time delay
In the linear case, a process with a pure time delay can be expressed in terms of a delayfree
part, plus a delay added at the output, as in Fig. 3.
Fig. 3. Process with time delay
The value
Pdelay
y at time k is measured, but not
P
y . In order to take into account the delay
in a control law formulation, prior knowledge of the delay value d is needed.
P
y can be
estimated as:
Delay free
Process
Delay, d
u y
P
y
Pdelay
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
139
( ) ( ) ( ) ( )
P Pdelay M M
y k y k y k y k d = + (53)
4.5 Tuning in PFC
According to the three principles of PFC, tuning is a function of the order of the basis
constructing the MV, the reference trajectory, the control horizon and the CLRT value.
The influence of the PFC parameters is given in Table 1, where the influence of various
PFC parameters is on precision (Steady State Resp.), transient response and robustness
are graded between 0 (indicating minimum influence) and 1 (indicating maximum
influence).
SS Resp. Transient Resp. Robustness
Basis function 1 0 0
Reference trajectory 0 1 1/2
Coincidence horizon 0 1/2 1
Table 1. Effect of PFC parameters in tuning
In most cases, an exponential reference trajectory is chosen along with a single coincidence
horizon point (H = 1) and a zero order basis function (Richalet, 1993). Considering the
known Open Loop Response Time of the system (OLRT), one can choose the CLRT value
given by the ratio OLRT/CLRT. This ratio then becomes the major tuning parameter
shaping the system output and MV, dictating how much overshoot occurs and ensuring
stability, on the condition that the internal model is accurate enough. For slow processes,
e.g., heat exchange systems, a ratio of 4 or 5 is found most suitable, and ensures a stable
PFC.
5. Simulation
Consider the single link manipulator with flexible joint shown in Figure 1. Choosing
1
q and
2
q as generalized coordinates, the kinetic energy is
2 2
1 2
1 1
2 2
K Iq Jq = + (54)
The potential energy is
2
1 1 2
1
(1 cos ) ( )
2
V MgL q k q q = + (55)
The Lagrangian is
2 2 2
1 2 1 1 2
1 1 1
(1 cos ) ( )
2 2 2
L K V Iq Jq MgL q k q q = = + (56)
Therefore we compute
Applications of Nonlinear Control
140
1
1
L
Iq
q
c
=
c
2
2
L
Jq
q
c
=
c
(57)
1
1
d L
Iq
dt q
c
=
c
2
2
d L
Jq
dt q
c
=
c
(58)
1 1 2
1
sin( ) ( )
L
MgL q k q q
q
c
=
c
1 2
2
( )
L
k q q
q
c
=
c
(59)
Therefore the equations of motion, ignoring damping, are given by
1 1 1 2
sin( ) ( ) 0 Iq MgL q k q q + + =
(60)
2 1 2
( ) Jq k q q u =
(61)
Note that since the nonlinearity enters into the first equation the control u cannot simply be
chosen to cancel it as in the case of the rigid manipulator equations. In other words, there is
no obvious analogue of the inverse dynamics control for the system in this form.
In state space we set
1 1
x q =
2 1
x q =
3 2
x q =
4 2
x q =
and write the system (60) (61) as
1 2
2 1 1 3
3 4
4 1 3
sin( ) ( )
1
( )
x x
MgL k
x x x x
I I
x x
k
x x x u
J J
=
=
=
= +
(62)
The system is thus of the form (15) with
2
1 1 3
4
1 3
sin( ) ( )
( )
( )
x
MgL k
x x x
I I
f x
x
k
x x
J
(
(
(
(
=
(
(
(
(
0
0
( ) 0
1
g x
J
(
(
(
( =
(
(
(
(63)
Therefore n=4 and the necessary and sufficient conditions for feedback linearization of this
system are that
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
141
{ }
2 3
2
2
0 0 0
0 0 0
, ( ), ( ), ( ) 4
1
0 0
1
0 0
f f f
k
IJ
k
IJ
rank g ad g ad g ad g rank
k
J J
k
J J
(
(
(
(
(
(
= =
(
(
(
(
(
(64)
which has rank 4 for 0 k > , , I J < . Also, since vector fields
{ }
2
, ( ), ( )
f f
g ad g ad g are constant,
they form an involutive set.
{ }
2
, ( ), ( )
f f
g ad g ad g (65)
To see this it suffices to note that the Lie Bracket of two constant vector fields is zero. Hence
the Lie Bracket of any two members of the set of vector fields in (65) is zero which is trivially
a linear combination of the vector fields themselves. It follows that the system (60) (61) is
feedback linearizable. The new coordinates
i i
y T = 1, , 4 i = (66)
are found from the conditions (67) (68)
1
, ( ) 0
k
f
dT ad g < >= 0, 1, , 2 k n = (67)
1
1
, ( ) 0
n
f
dT ad g
< >=
(68)
with n=4, that is
1
, 0 dT g < >= (69)
1
,[ , ] 0 dT f g < >= (70)
2
1
, ( ) 0
f
dT ad g < >= (71)
3
1
, ( ) 0
f
dT ad g < >= (72)
Carrying out the above calculations leads to the system of equations
1
2
0
T
x
c
=
c
1
3
0
T
x
c
=
c
1
4
0
T
x
c
=
c
(73)
and
1
1
0
T
x
c
=
c
(74)
Applications of Nonlinear Control
142
From this we see that the function
1
T should be a function of
1
x alone. Therefore, we take
the simplest solution
1 1 1
y T x = = (75)
and compute
2 2 1 2
, y T dT f x = =< >= (76)
3 3 2 1 1 3
, sin( ) ( )
MgL k
y T dT f x x x
I I
= =< >= (77)
4 4 3 1 2 2 4
, cos( ) ( )
MgL k
y T dT f x x x x
I I
= =< >=
(78)
The feedback linearizing control input u is found from the condition
4
4
1
( , )
,
( ( )) ( ) ( )
u v dT f
dT g
IJ
v a x x v x
k
 o
= < >
< >
= = +
(79)
where
2
1 2 1 1 3 1
( ) : sin( )( cos( ) ) ( )( cos( ))
MgL MgL MgL k k k k
x x x x x x x
I I I I I J I
o = + + + + +
(80)
Therefore in the coordinates
1 4
, , y y with the control law (79) the system becomes
1 2
y y =
2 3
y y =
3 4
y y =
4
y v =
or, in matrix form,
y Ay bv = + (81)
where
0 1 0 0
0 0 1 0
0 0 0 1
0 0 0 0
A
(
(
(
=
(
(
0
0
0
1
b
(
(
(
=
(
(
The transformed variables
1 4
, , y y are themselves physically meaningful. We see that
1 1
y x = =link position
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
143
2 2
y x = =link velocity
3 2
y y = =link acceleration
4 3
y y = =link jerk
Since the motion trajectory of the link is typically specified in terms of these quantities they
are natural variables to use for feedback.
For given a linear system in state space form, such as (81), a state feedback control law is an
input v of the form
4
1
T
i i
i
v k y r k y r
=
= + = +
(82)
where
i
k are constants and r is a reference input. If we substitute the control law (82) into
(81), we obtain
( )
T
y A bk y br = + (83)
Thus we see that the linear feedback control has the effect of changing the poles of the
system from those determined by A to those determined by
T
A bk
When the parameters are chosen
1
62.5 k = ,
2
213.6 k = ,
3
204.2 k =
4
54 k = , we can get step
responses in Figure 4. where k1, k2, k3 and k4 are linear feedback coefficients to place the
eigenvalues of A in a desired location.
1 1
2 2
3 3
4 4
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
62.5 213.8 204.2 54 1
y y
y y
r
y y
y y
( ( ( (
( ( ( (
( ( ( (
= +
( ( ( (
( ( ( (
(84)
 
1
2
3
4
1 0 0 0
y
y
y
y
y
(
(
(
=
(
(
(85)
The internal model parameter: 0.016
M
K = , 3
M
T = , 8 d = , and the coincidence point H=10.
Response time of reference trajectory is 0.01, and sample time is 0.01.
For the uncertainty q , the system (38) can be written in matrix form as
( ) { ( , )}
T
y A bk y b r y v q = + +
then use predictive function control strategy to reduce or overcome uncertainty of nonlinear
feedback error ( , ) y v q ,and simulation result is shown in Figure 5 for ( , ) 10%
r
y v y q = .
Applications of Nonlinear Control
144
Fig. 4. link position output
Predictive Function Control of the SingleLink Manipulator with Flexible Joint
145
Fig. 5. link position output with uncertainty rejection
Applications of Nonlinear Control
146
6. Conclusion
A new three stage design method is presented for the single link manipulator with flexible
joint. The first is feedback linearization; the second is to use pole placement to satisfy
performance, and the third is to develop predictive function control to compensate
uncertainty. Finally, for the same uncertainty, robustness is better than traditional method.
7. References
Khorasani, K. (1990). Nonlinear feedback control of flexible joint manipulators: a single link
case study. IEEE Transactions on Automatic Control, Vol. 35, No. 10, pp. 11451149,
ISSN 00189286
Richalet, J.; Rault, A.; Testud, J. L.; Papon, J.(1978). Model predictive heuristic control:
Applications to industrial processes. Automatica, Vol. 14, No. 5, pp. 413428, ISSN
00051098
Spong, M.W.; Khorasani, K. & Kokotovic P.V. (1987). An integral manifold approach to the
feedback control of flexible joint robots. IEEE Journal of Robotics and Automation, Vol.
RA3, No. 4, pp. 291300, ISSN 08824967
Spong, M.W.(1987). Modeling and control of elastic joint robots. Journal of Dyn. Sys. Meas.
and Cont., Vol. 109, pp.310319, ISSN 00220434
Spong, M.W. (1989). On the force control problem for flexible joint manipulator. IEEE
Transactions on Automatic Control, Vol. 34, No. 1, pp. 107111, ISSN 00189286
E. F. Camacho, Carlos Bordons(2004). Model predictive control. SpringerVerlag London Berlin
Limited. ISBN 1852336943
1. Introduction
Several classes of general hybrid and switched dynamic systems have been extensively
studied, both in theory and practice [3,4,7,11,14,17,19,26,27,30]. In particular, driven by
engineering requirements, there has been increasing interest in optimal design for hybrid
control systems [3,4,7,8,13,17,23,26,27]. In this paper, we investigate some specic types of
hybrid systems, namely hybrid systems of mechanical nature, and study the corresponding
hybrid OCPs. The class of dynamic models to be discussed in this work concerns hybrid
systems where discrete transitions are being triggered by the continuous dynamics. The
control objective (control design) is to minimize a cost functional, where the control
parameters are the conventional control inputs.
Recently, there has been considerable effort to develop theoretical and computational
frameworks for complex control problems. Of particular importance is the ability to operate
such systems in an optimal manner. In many realworld applications a controlled mechanical
system presents the main modeling framework and can be specied as a strongly nonlinear
dynamic system of high order [9,10,22]. Moreover, the majority of applied OCPs governed
by sophisticated realworld mechanical systems are optimization problems of the hybrid
nature. The most realworld mechanical control problems are becoming too complex to allow
analytical solution. Thus, computational algorithms are inevitable in solving these problems.
There is a number of results scattered in the literature on numerical methods for optimal
control problems. One can nd a fairly complete review in [3,4,8,24,25,29]. The main idea
of our investigations is to use the variational structure of the solution to the specic two point
boundaryvalue problem for the controllable hybridtype mechanical systems in the form of
EulerLagrange or Hamilton equation and to propose a new computational algorithm for the
associated OCP. We consider an OCP in mechanics in a general setting and reduce the initial
problem to a constrained multiobjective programming. This auxiliary optimization approach
provides a basis for a possible numerical treatment of the original problem.
The outline of our paper is as follows. Section 2 contains some necessary basic facts
related to the conventional and hybrid mechanical models. In Section 3 we formulate
and study our main optimization problem for hybrid mechanical systems. Section 4 deals
with the variational analysis of the OCP under consideration. We also briey discuss the
On Optimization Techniques for a Class
of Hybrid Mechanical Systems
Vadim Azhmyakov and Arturo Enrique Gil Garca
Department of Control Automation, CINVESTAV, A.P. 14740, Mexico D.F.,
Mexico
9
2 WillbesetbyINTECH
computational aspect of the proposed approach. In Section 5 we study a numerical example
that constitutes an implementable hybrid mechanical system. Section 6 summarizes our
contribution.
2. Preliminaries and some basic facts
Let us consider the following variational problem for a hybrid mechanical system that is
characterized by a family of Lagrange functions {
L
p
i
}, p
i
P
minimize
_
1
0
r
i=1
[t
i1
,t
i
)
(t)
L
p
i
(t, q(t), q(t))dt
subject to q(0) = c
0
, q(1) = c
1
,
(1)
where P is a nite set of indices (locations) and q() (q(t) R
n
) is a continuously differentiable
function. Here
[t
i1
,t
i
)
() are characteristic functions of the time intervals [t
i1
, t
i
), i = 1, ..., r
associated with locations. Note that a full time interval [0, 1] is assumed to be separated into
disjunct subintervals of the above type for a sequence of switching times:
:= {t
0
= 0, t
1
, ..., t
r
= 1}.
We refer to [3,4,7,8,13,17,23,26,27] for some concrete examples of hybrid systems with the
above dynamic structure. Consider a class of hybrid mechanical systems that can be
represented by n generalized conguration coordinates q
1
, ..., q
n
. The components q
(t), =
1, ..., n of q(t) are the socalled generalized velocities. Moreover, we assume that
L
p
i
(t, , )
are twice continuously differentiable convex functions. It is well known that the formal
necessary optimality conditions for the given variational problem(1) describe the dynamics of
the mechanical systemunder consideration. This description can be given for every particular
location and nally, for the complete hybrid system. In this contribution, we study the hybrid
dynamic models that free from the possible external inuences (uncertainties) or forces. The
optimality conditions for mentioned above can be rewritten in the form of the secondorder
EulerLagrange equations (see [1])
d
dt
L
p
i
(t, q, q)
q
L
p
i
(t, q, q)
q
= 0, = 1, ..., n ,
q(0) = c
0
, q(1) = c
1
,
(2)
for all p
i
P. The celebrated Hamilton Principle (see e.g., [1]) gives an equivalent variational
characterization of the solution to the twopoint boundaryvalue problem (2).
For the controllable hybrid mechanical systems with the parametrized (control inputs)
Lagrangians L
p
i
(t, q, q, u), p
i
P we also can introduce the corresponding equations of
motion
d
dt
L
p
i
(t, q, q, u)
q
L
p
i
(t, q, q, u)
q
= 0,
q(0) = c
0
, q(1) = c
1
,
(3)
148 Applications of Nonlinear Control
On Optimization Techniques for a Class of Hybrid Mechanical Systems 3
where u() U is a control function from the set of admissible controls U. Let
U := {u R
m
: b
1,
u
b
2,
, = 1, ..., m},
U := {v() L
2
m
([0, 1]) : v(t) U a.e. on [0, 1]},
where b
1,
, b
2,
, = 1, ..., m are constants. The introduced set U provides a standard example
of an admissible control set. In this specic case we deal with the following set of admissible
controls U
C
1
m
(0, 1). Note that L
p
i
depends directly on the control function u(). Let us
assume that functions L
p
i
(t, , , u) are twice continuously differentiable functions and every
L
p
i
(t, q, q, ) is a continuously differentiable function. For a xed admissible control u() we
obtain for all p
i
P the above hybrid mechanical system with
L
p
i
(t, q, q) L
p
i
(t, q, q, u(t)).
It is also assumed that L
p
i
(t, q, , u) are strongly convex functions, i.e., for any
(t, q, q, u) RR
n
R
n
R
m
, R
n
the following inequality
n
,=1
2
L
p
i
(t, q, q, u)
q
=1
, > 0
holds for all p
i
P. This natural convexity condition is a direct consequence of the
classical representation for the kinetic energy of a conventional mechanical system. Under
the abovementioned assumptions, the twopoint boundaryvalue problem (3) has a solution
for every admissible control u() U [18]. We assume that (3) has a unique solution for every
u() U. For an admissible control u() U the solution to the boundaryvalue problem (3)
is denoted by q
u
(). We call (3) the hybrid EulerLagrange control system. Let us now give a
simple example of the above mechanical model.
Example 1. We consider a variable linear massspring system attached to a moving frame that is a
generalization of the corresponding system from [22]. The considered control u() U
C
1
1
(0, 1) is
the velocity of the frame. By
p
i
we denote the variable masses of the system. The kinetic energy
K = 0.5
p
i
( q + u)
2
depends on the control input u(). Therefore, we have
L
p
i
(q, q, u) = 0.5(
p
i
( q + u)
2
q
2
), R
+
and
d
dt
L
p
i
(t, q, q, u)
q
L
p
i
(t, q, q, u)
q
=
p
i
( q + u) + q = 0.
By we denote here the elasticity coefcient of the spring system.
Note that some important controlled mechanical systems have a Lagrangian function of the
following form (see e.g., [22])
L
p
i
(t, q, q, u) = L
0
p
i
(t, q, q) +
m
=1
q
.
149 On Optimization Techniques for a Class of Hybrid Mechanical Systems
4 WillbesetbyINTECH
In this special case we easily obtain
d
dt
L
0
p
i
(t, q, q)
q
L
0
p
i
(t, q, q)
q
=
_
u
= 1, ..., m,
0 = m +1, ..., n.
Note that the control function u() is interpreted here as an external force.
Let us now consider the Hamiltonian reformulation for the controllable EulerLagrange
system (3). For every location p
i
fromP we introduce the generalized momenta
s
:= L
p
i
(t, q, q, u)/ q
=1
s
L
p
i
(t, q, q, u).
In the case of hyperregular Lagrangians L
p
i
(t, q, q, u) (see e.g., [1]) the Legendre transform,
namely, the mapping
L
p
i
: (t, q, q, u) (t, q, s, u),
is a diffeomorphism for every p
i
P,. Using the introduced Hamiltonian H(t, q, s, u), we can
rewrite system (3) in the following Hamiltontype form
q
(t) =
H
p
i
(t, q, s, u)
s
, q(0) = c
0
, q(1) = c
1
,
s
(t) =
H
p
i
(t, q, s, u)
q
, = 1, ..., n .
(4)
Under the abovementioned assumptions, the boundaryvalue problem (4) has a solution for
every u() U. We will call (4) a Hamilton control system. The main advantage of (4) in
comparison with (3) is that (4) immediately constitutes a control system in standard state
space form with state variables (q, s) (in physics usually called the phase variables). Consider
the system of Example 1 with
H
p
i
(q, s, u) =
1
2
p
i
( q
2
u
2
) +
1
2
q
2
su.
The Hamilton equations in this case are given as follows
q =
H
p
i
(q, s, u)
s
=
1
p
i
s u,
s =
H
p
i
(q, s, u)
q
= q.
Clearly, for
L
p
i
(t, q, q, u) = L
0
p
i
(t, q, q) +
m
=1
q
, u
=1
q
,
where H
0
p
i
(t, q, s) is the Legendre transform of L
0
p
i
(t, q, q).
3. Optimization of control processes in hybrid mechanical systems
Let us formally introduce the class of OCPs investigated in this paper:
minimize J :=
_
1
0
r
i=1
[t
i1
,t
i
)
(t) f
0
p
i
(q
u
(t), u(t))dt
subject to u(t) U t [0, 1], t
i
, i = 1, ..., r,
(5)
where f
0
p
i
: [0, 1] R
n
R
m
R be continuous and convex on R
n
R
m
objective functions.
We have assumed that the boundaryvalue problems (3) have a unique solution q
u
() and that
the optimization problem (5) also has a solution. Let (q
opt
(), u
opt
()) be an optimal solution
of (5). Note that we can also use the associated Hamiltoniantype representation of the initial
OCP (5). We mainly focus our attention on the application of direct numerical algorithms to
the hybrid optimization problem (5). A great amount of works is devoted to the direct or
indirect numerical methods for conventional and hybrid OC problems (see e.g., [8,24,25,29]
and references therein). Evidently, an OC problem involving ordinary differential equations
can be formulated in various ways as an optimization problem in a suitable function space
and solved by some standard numerical algorithms (e.g., by applying a rstorder methods
[3,24]).
Example 2. Using the EulerLagrange control system from Example 1, we now examine the
following OCP
minimize J :=
_
1
0
r
i=1
[t
i1
,t
i
)
(t)k
p
i
(u(t) + q(t))dt
subject to q(t) +
p
i
q(t) = u(t), i = 1, ..., r,
q(0) = 0, q(1) = 1,
u() C
1
1
(0, 1), 0 u(t) 1 t [0, 1],
where k
p
i
are given (variable) coefcients. Let
p
i
4/
2
for every p
i
P. The solution q
u
() of
the associated boundaryvalue problem can be written as follows
q
u
(t) = C
u
i
sin (t
_
/
p
i
)
_
t
0
_
/
p
i
sin (
_
/
p
i
(t l)) u(l)dl,
where t [t
i1
, t
i
), i = 1, ..., r and
C
u
i
=
1
sin
_
/
p
i
[1 +
_
1
0
_
/
p
i
sin (
i=1
[t
i1
,t
i
)
(t)k
p
i
[u(t) + q
u
(t)]dt =
_
1
0
r
i=1
[t
i1
,t
i
)
(t)k
p
i
[u(t) + C
u
i
sin (t
_
/
p
i
)
_
t
0
_
/
p
i
sin (
_
/
p
i
(t l)) u(l)dl]dt.
Let now k
p
i
= 1 for all p
i
P. Using the Hybrid Maximum Principle (see [4]), we conclude that
the admissible control u
opt
(t) 0.5 is an optimal solution of the given OCP. Note that this result is
also consistent with the Bauer Maximum Principle (see e.g., [2] ). For u
opt
() we can compute the
corresponding optimal trajectory given as follows
q
opt
(t) =
sin (t
_
/
p
i
)
sin
_
/
p
i
, t [t
i1
, t
i
), i = 1, ..., r.
Evidently, we have
_
/
p
i
/2 for every location p
i
. Moreover, q
opt
(t) 3 under the condition
q
opt
() C
1
1
(0, 1).
Finally, note that a wide family of classical impulsive control systems (see e.g., [12]) can be
described by the conventional controllable EulerLagrange or Hamilton equations (see [5]).
Moreover, we refer to [6] for impulsive hybrid control systems and associated OCPs. Thus the
impulsive hybrid systems of mechanical nature can also be incorporated into the modeling
framework presented in this section.
4. The variational approach to hybrid OCPs of mechanical nature
An effective numerical procedure, as a rule, uses the specic structure of the problem under
consideration. Our aim is to study the variational structure of the main OCP (5). Let
i
:= {() C
1
n
([t
i1
, t
i
]) : (t
i1
) = c
i1
, (t
i
) = c
i
},
where i = 1, ..., r.. The vectors c
i
, where i = 1, ..., r are dened by the corresponding switching
mechanism of a concrete hybrid system. We refer to [3,4,26] for some possible switching rules
determined for various classes of hybrid control systems. We now present an immediate
consequence of the classical Hamilton Principle from analytical mechanics.
Theorem 1. Let all Lagrangians L
p
i
(t, q, q, u) be a strongly convex function with respect to the
generalized variables q
i
, i = 1, ..., n. Assume that every boundaryvalue problem from (3) has a unique
solution for every u() U
C
1
m
(0, 1). A function q
u
(), where u() U
C
1
m
(0, 1), is a solution
of the sequence of boundaryvalue problems (3) if and only if a restriction of this function on every time
interval [t
i1
t
i
), i = 1, ..., r can be found as follows
q
u
i
() = argmin
q()
i
_
t
i
t
i1
L
p
i
(t, q(t), q(t), u(t))dt.
152 Applications of Nonlinear Control
On Optimization Techniques for a Class of Hybrid Mechanical Systems 7
For an admissible control function u() from U we now introduce the following two
functionals
T
p
i
(q(), z()) :=
_
t
i
t
i1
[L
p
i
(t, q(t), q(t), u(t)) L
p
i
(t, z(t), z(t), u(t))]dt,
V
p
i
(q()) := max
z()
i
_
t
i
t
i1
[L
p
i
(t, q(t), q(t), u(t)) L
p
i
(t, z(t), z(t), u(t))]dt,
for all indexes p
i
P. Generally, we dene q
u
() as an element of the Sobolev space
W
1,
n
(0, 1), i.e., the space of absolutely continuous functions with essentially bounded
derivatives. Let us give a variational interpretation of the admissible solutions q
u
() to a
sequence of problems (3).
Theorem 2. Let all Lagrangians L
p
i
(t, q, q, u) be strongly convex functions with respect to the
variables q
i
, i = 1, ..., n. Assume that every boundaryvalue problem from (3) has a unique solution for
every u() U
C
1
m
(0, 1). An absolutely continuous function q
u
(), where u() U
C
1
m
(0, 1),
is a solution of the sequence of problems (3) if and only if a restriction of this function on
[t
i1
t
i
), i = 1, ..., r can be found as follows
q
u
i
() = argmin
q()W
1,
n
(t
i1
,t
i
)
V
p
i
(q()) (6)
Proof. Let q
u
() W
1,
n
(t
i1
, t
i
) be a unique solution of a partial problem (3) on the
corresponding time interval, where u() U
C
1
m
(0, 1). Using the Hamilton Principle in
every location p
i
inP, we obtain the following relations
min
q()W
1,
n
(t
i1
,t
i
)
V
p
i
(q()) = min
q()W
1,
n
(t
i1
,t
i
)
max
z()
i
_
t
i
t
i1
[L
p
i
(t, q(t), q(t), u(t))
_
t
i
t
i1
L
p
i
(t, z(t), z(t), u(t))]dt = min
q()W
1,
n
(t
i1
,t
i
)
_
t
i
t
i1
L
p
i
(t, q(t), q(t), u(t))dt
min
z()
i
_
t
i
t
i1
L
p
i
(t, z(t), z(t), u(t))dt =
_
t
i
t
i1
L
p
i
(t, q
u
(t), q
u
(t), u(t))dt
_
t
i
t
i1
L
p
i
(t, q
u
(t), q
u
(t), u(t))dt = V
p
i
(q
u
()) = 0.
If the condition (6) is satised, then q
u
() is a solution of the sequence of the boundaryvalue
problem (3). This completes the proof.
Theorem 1 and Theorem 2 make it possible to express the initial OCP (5) as a multiobjective
optimization problem over the set of admissible controls and generalized coordinates
minimize J(q(), u()) and P(q())
subject to
(q(), u()) (
_
i=1,...,r
i
) (U
C
1
m
(0, 1)),
(7)
153 On Optimization Techniques for a Class of Hybrid Mechanical Systems
8 WillbesetbyINTECH
or
minimize J(q(), u()) and V(q())
subject to
(q(), u()) (
_
i=1,...,r
i
) (U
C
1
m
(0, 1)),
(8)
where
P(q()) :=
_
1
0
r
i=1
[t
i1
,t
i
)
(t)L
p
i
(t, q(t), q(t), u
opt
(t))dt
and
V(q()) :=
[t
i1
,t
i
)
(t)V
p
i
(q()).
The auxiliary minimizing problems (7) and (8) are multiobjective optimization problems (see
e.g., [16,28]). Note that the set
(U
C
1
m
(0, 1)
is a convex set. Since f
0
(t, , ), t [0, 1] is a convex function, J(q(), u()) is also convex. If
P() (or V()) is a convex functional, then we deal with a convex multiobjective minimization
problem (7) (or (8)).
The variational representation of the solution of the twopoint boundaryvalue problem (3)
eliminates the differential equations from the consideration. The minimization problems
(7) and (8) provide a basis for numerical algorithms to the initial OCP (5). The auxiliary
optimization problem (7) has two objective functionals. For (7) we now introduce the
Lagrange function [28]
(t, q(), u(), ,
3
) :=
1
J(q(), u()) +
2
P(q())+
3
dist
(
i=1,...,r
i
)(U
C
1
m
(0,1))
{(q(), u())},
where dist
(
i=1,...,r
i
)(U
C
1
m
(0,1))
{} denotes the distance function
dist
(
i
)(U
C
1
m
(0,1))
{(q(), u())} := inf{(q(), u())

C
1
n
(0,1)C
1
m
(0,1)
, (
_
i=1,...,r
i
) (U
C
1
m
(0, 1))}.
We also used the following notation
:= (
1
,
2
)
T
R
2
+
.
Note that the above distance function is associated with the Cartesian product
(
_
i=1,...,r
i
) (U
C
1
m
(0, 1)).
154 Applications of Nonlinear Control
On Optimization Techniques for a Class of Hybrid Mechanical Systems 9
Recall that a feasible point (q
(), u
(), u
()).
A necessary condition for (q
(), u
R
2
+
such that
0
(q(),u())
(t, q
(), u
(),
,
3
). (9)
By
(q(),u())
we denote here the generalized gradient of the Lagrange function . We refer to
[28] for further theoretical details. If P() is a convex functional, then the necessary condition
(9) is also sufcient for (q
(), u
(), u
N
i
) (U
N
C
1
m,N
(0, 1)),
(10)
where J
N
and P
N
are discrete variants of the objective functionals J and P from (7). Moreover,
N
i
is a correspondingly discrete set
i
and C
1
m,N
(0, 1) is set of suitable discrete functions
that approximate the trajectories set C
1
m
(0, 1). Note that the initial continuous optimization
problem can also be presented in a similar discrete manner. For example, we can introduce
the (Euclidean) spaces of piecewise constant trajectories q
N
() and piecewise constant control
functions u
N
(). As we can see the Banach space C
1
n
(0, 1) and the Hilbert space L
2
m
([0, 1]) will
be replaced in that case by some appropriate nitedimensional spaces.
The discrete optimization problem (10) approximates the innitedimensional optimization
problem (7). We assume that the set of all weak Pareto optimal solution of the discrete
problem (10) is nonempty. Moreover, similarly to the initial optimization problem (7) we
also assume that the discrete problem (10) is regular. If P() is a convex functional, then the
discrete multiobjective optimization problem (10) is also a convex problem. Analogously to
the continuous case (7) or (8) we also can write the corresponding KKT optimality conditions
for a nitedimensional optimization problem over the set of variables (q
N
(), u
N
()). The
necessary optimality conditions for a discretized problem (10) reduce the nitedimensional
multiobjective optimization problem to a system of nonlinear equations. This problem can be
solved by some gradientbased or Newtontype methods (see e.g., [24]).
Finally, note that the proposed numerical approach uses the necessary optimality conditions,
namely the KKT conditions, for the discrete variant (10) of the initial optimization problem(7).
It is common knowledge that some necessary conditions of optimality for discrete systems, for
example the discrete version of the classical Pontryagin Maximum Principle, are noncorrect
in the absence of some restrictive assumptions. For a constructive numerical treatment of
the discrete optimization problem it is necessary to apply some suitable modications of
the conventional optimality conditions. For instance, in the case of discrete optimal control
problems one can use socalled Approximate Maximum Principle which is specially designed
for discrete approximations of general OCPs [21].
5. Mechanical example
This section is devoted to a short numerical illustration of the proposed hybrid approach
to mechanical systems. We deal with a practically motivated model that has the following
structure (see Fig. 1).
Let us rstly describe the parameters of the mechanical model under consideration:
156 Applications of Nonlinear Control
On Optimization Techniques for a Class of Hybrid Mechanical Systems 11
Fig. 1. Mechanical example
q
1
it corresponds to the position of motor.
q
2
is the position of inertia J
2
.
J
1
, J
2
are the external inertias.
J
m
is an inertia of motor.
B
m
it corresponds to the friction of the motor.
B
1
, B
2
they correspond to the frictions of the inertias J
1
, J
2
.
k is a constant called the rate or spring constant.
u it corresponds to the torque of motor.
The relations for the kinetic potential energies give a rise to the corresponding Lagrange
dynamics:
K(t) =
1
2
J
m
q
2
1
+
1
2
J
2
q
2
2
V(t) =
1
2
k (q
1
q
2
)
2
Finally, we have
L(q(t), q(t)) =
1
2
J
m
q
2
1
+
1
2
J
2
q
2
2
1
2
k (q
1
q
2
)
2
and the EulerLagrange equation with respect to the generalized coordinate q
1
has the
following form
J
m
q
1
+ B
m
q
1
k(q
2
(t) q
1
(t)) = u(t) (11)
We now considered the EulerLagrange equation with respect to the second generalized
variable, namely, with respect to q
2
d
dt
L(q(t), q(t))
q
2
L(q(t), q(t))
q
2
= B
2
q
2
(t)
We get the next relation
J
2
q
2
(t) + B
2
q
2
(t) + k(q
2
(t) q
1
(t)) = 0
157 On Optimization Techniques for a Class of Hybrid Mechanical Systems
12 WillbesetbyINTECH
The redenition of the states x
1
:= q
1
, x
2
:= q
1
, x
3
:= q
2
, x
4
:= q
2
with X := (x
1
, x
2
, x
3
, x
4
)
T
implies the compact statespace form of the resulting equation:
X :=
x
1
x
2
x
3
x
4
0 1 0 0
k
J
m
B
m
J
m
k
J
m
0
0 0 0 1
k
J
2
0
k
J
2
B
2
J
2
x
1
x
2
x
3
x
4
0
1
J
m
0
0
u, X
0
:=
x
0
1
x
0
2
x
0
3
x
0
4
(12)
The switching structure of the system under consideration is characterized by an additional
inertia J
1
and the associated friction B
1
. The modied energies are given by the expressions:
the kinetic energy:
K(t) =
1
2
J
m
q
2
1
+
1
2
J
1
q
2
1
+
1
2
J
2
q
2
2
the potential energy:
V(t) =
1
2
k (q
1
q
2
)
2
The function of Lagrange can be evaluated as follows
L(q, q) =
1
2
J
m
q
2
1
+
1
2
J
1
q
2
1
+
1
2
J
2
q
2
2
1
2
k (q
1
q
2
)
2
(13)
The resulting EulerLagrange equations (with respect to q
1
and to q
2
can be rewritten as
(J
m
+ J
1
) q
1
(t) + (B
m
+ B
1
) q
1
(t) k(q
2
(t) q
1
(t)) = u(t)
J
2
q
2
(t) + B
2
q
2
(t) + k(q
2
(t) q
1
(t)) = 0
(14)
Using the notation introduced above, we obtain the nal statespace representation of the
hybrid dynamics associated with the given mechanical model:
X :=
x
1
x
2
x
3
x
4
0 1 0 0
k
J
m
+J
1
(B
m
+B
1
)
J
m
+J
1
k
J
m
+J
1
0
0 0 0 1
k
J
2
0
k
J
2
B
2
J
2
x
1
x
2
x
3
x
4
0
1
J
m
+J
1
0
0
u (15)
The considered mechanical system has a switched nature with a statedependent switching
signal. We put x
4
= 10 for the switchinglevel related to the additional inertia in the system
(see above).
Our aim is to nd an admissible control law that minimize the value of the quadratic costs
functional
I(u()) =
1
2
_
t
f
t
0
_
X
T
(t)QX(t) + u
T
(t)Ru(t)
_
dt min
u()
(16)
The resulting Linear Quadratic Regulator that has the follow form
u
opt
(t) = R
1
(t)B
T
(t)P(t)X
opt
(t) (17)
158 Applications of Nonlinear Control
On Optimization Techniques for a Class of Hybrid Mechanical Systems 13
where P(t) is a solution of the Riccati equation (see [7] for details)
P(t) = (A
T
(t)P(t) + P(t)A(t)) + P(t)B(t)R
1
(t)B
T
(t)P(t) Q(t) (18)
with the nal condition
P(t
f
) = 0 (19)
Let us now present a conceptual algorithm for a concrete computation of the optimal pair
(u
opt
, X
opt
()) in this mechanical example. We refer to [7, 8] for the necessary facts and the
general mathematical tool related to the hybrid LQtechniques.
Algorithm 1. The conceptual algorithm used:
(0) Select a t
swi
_
0, t
f
_
, put an index j = 0
(1) Solve the Riccati euqation (18) for (15) on the time intervals [0, t
swi
]
_
t
swi
, t
f
_
(2) solve the initial problem (12) for (17)
(3) calculate x
4
(t
swi
) + 10, if  x
4
(t
swi
) + 10 
(1)
where the state x e R
n
, the input u e R
m
, the output y e R
p
, and F and H are vectorvalued
functions with F : RR
n
R
m
R
n
and H : RR
n
R
m
R
p
.
Applications of Nonlinear Control
164
A second kind of used model is the discrete dynamic system defined by a finite set of
difference equations:
 
 
0 0
( 1) , ( ), ( ) ( )
( ) , ( ), ( )
x k F k x k u k x k x
y k H k x k u k
+ = =
=
(2)
where x(k) = x(kh), u(k) = u(kh), h is the sampling time interval, and k 0 is an integer.
The objective of this chapter is to propose a new strategy for control design using
optimization method which is suitable for real time applications. This new methodology is
based on neural network which is the classical approach to treat practical results using
experimental tests. In order to illustrate this methodology and its applications, we will
present an example of the intake air manifold control in a Diesel internal combustion engine.
The chapter is divided as follows: In the second section a short overview of classical control
methods is presented. In the third section a new methodology for control is proposed. In the
fourth section, we present the application of the new control methodology to the Diesel
engine. And finally, we end this chapter with our conclusions and remarks.
2. Overview of classical control methods
A main goal of the feedback control system is to guarantee the stability of the closedloop
behavior. For linear systems, this can be obtained by adapting the control parameters of the
transfer function which describes the system in a way so that the real parts of its poles have
negative values. Otherwise, Nonlinear control systems use specific theories to ensure the
system stability and that is regardless the inner dynamic of the system. The possibility to
realize different specifications varies according to the model considered and the control
strategy chosen. Hereafter we present a summary of some techniques that can be used:
2.1 Theory of Lyapunov
Lyapunov theory is usually used to determine the stability properties at an equilibrium
point without the need to resolve the state equations of the nonlinear system. Let us
consider the autonomous nonlinear system
( ) x F x = (3)
where x e R
n
is the state variable of the system and F is a smooth vector field. Assume that
there is a function V defined as follows:
:
n
V R R
+
so that ( ) 0 0 V x x = = and lim ( )
x
V x
(6)
is reduced to the origin.
These two theorems are the base of a large number of results on analysis of stability for
nonlinear systems. In fact, the theory of Lyapunov KrosoviskiiLaSalle is fundamental and
is the base of this analysis. In the literature, this theory can have various versions according
to the nature of the problem, for instance, for discrete models, stochastic systems or partial
differential equations.
In addition to the methodologies developed before, the theory is used to describe the control
problems. The use of this theory is illustrated by the following result of feedback stabilization.
Let us consider the following controlled system
( ) ( ) x F x u G x = +
(7)
where
n
x R e is the state, u R e is the control variable, F and G are smooth vector fields.
Assume there is V a Lyapunov function so that
( ), ( ) 0 V V x F x = V s
(8)
Under some hypothesis is proved (Outbib, 1999) that the closedloop system defined from
(7) with
( ), ( ) u V x G x = V
(9)
is globally asymptotically stable at the origin. A simple example to illustrate this result is the
scalar system
x u = (10)
Clearly, the system verifies the hypothesis with
2
( ) 1 /2 V x x = and the stabilizing
control u x = can be deduced. This approach has been applied to practical process (Outbib,
2000; Dreyfus, 1962)
2.2 Adaptive control
The adaptive control is mainly used in cases where the control law must be continuously
adapted due to the varying nature of the system parameters or its initial uncertainties.
Let us consider the following non linear system
( , ) x F x u =
(11)
Where x denotes the state variable of the system and u designates a parameter. The
adaptive control is used in the situation where the parameter u is not known or can change.
For example, let us consider the scalar classical system:
Applications of Nonlinear Control
166
u = +
2
x x u
(12)
If u is known the system (12) can be globally asymptotically stable using a control law of the
form
2
( ) u x k x u = , where k is any smooth scalar function defined as follow: ( ) 0 k x x > for
0 x = .
The certaintyequivalent controller is defined by
2
( )
u x k x
w
u
u
(13)
where w is the update law.
Let V be the Lyapunov function defined by:
( )
2
2
1
( , )
2 2
V x x = +
o
u u u (14)
with 0 o > . The derivative of the closedloop system defined from (12) and (13):
2
( ) ( )
u x k x
w
u u
u
(15)
is given by
( ) ( )
3
( , ) ( ) V x x k x x w u u u o u u = +
(16)
Now, if we let ( )
3
1 / w x o = , we get
( , ) ( ) 0 V x x k x for all x u = s
(17)
This implies that ( , ) x u
(19)
than the system can be globally asymptotically stable using appropriate values for
1
a
and
2
a .
2.4 Optimal control
The objective of the optimal control method is to search for the best dynamic course which is
capable of transporting the system from an initial state to a final desired state at minimum
cost. An example of its various applications can be found in the satellite control. More
precisely, the optimal control technique can be defined as follows:
Let us consider the following system:
( , ( ), ( )) x F t x t u t =
(20)
where x e R
n
designates the state variable and u e R
m
is the control variable.
:
n m n
f R R R R is a smooth vectorvalued function .The optimal control is to find a
suitable dynamic control u(t) which allows the system to follow an optimal trajectory ( ) x t
that minimizes the cost function :
1
0
( , ( ), ( ))
t
t
J H t x t u t =
}
(21)
Several approaches have been used to resolve this problem. Among these approaches we
can cite the variational calculus (Dreyfus, 1962), the maximum principle of Pontryagin
(Pontryagin, 1962) or the procedure of dynamic programming method of Bellman (Bellman,
1957).
Let us consider a simple example such as the hanging pendulum. The equation describing
the movement of the pendulum angular position under an applied torque o is given by:
2
1 2
( ) ( ) ( ) ( )
(0) (0)
t t t t u u e u o
u u u u
+ + =
= =
(22)
where u designates the angular position at time t . Let x be the system state variable
( ), ( ) x t t u u ( =
, we can write :
Applications of Nonlinear Control
168
2
2
2 1
( ) ( , )
x
x t f x
x x
o
e o
 
= =


+
\ .
(23)
Therefore the optimal control goal can be to minimize the time intervalt , in order to reach
the state values ( ) 0 x t = .
2.5 Robust control
The objective of robust control is to find a control law in such a way that the response of the
system and the error signals are maintained to desired values despite the effects of
uncertainties on the system. The uncertainties sources can be any disturbance signals, the
measurement noise or the modeling errors due to none considered nonlinearities and time
varying parameters.
The theory of robust control began in the 1970s and 1980s (Doyle, 1979; Zames, 1981) with
some aircraft applications. Actually, its applications concern different domains (aerospace,
economics, ...).
3. New algorithm for Optimized Nonlinear Control (ONC)
The objective of this methodology is to propose a system optimized dynamic control which
can be used in real time control applications. The proposed methodology (Omran, 2008b)
can be divided into five steps: 1) Modeling process, 2) Model validation, 3) Dynamic
optimization process, 4) Creation of a large database of the optimal control variables using
the dynamic optimization process, 5) The neural network controller.
In the next subsections, we present the different methodology steps and we explain its
application using the example of the Diesel engine system.
3.1 Modeling process
The general equations which describe the functioning of a system can be expressed using
the following form (Rivals, 1995):
( , , , )
( , , , )
X F X I u t
Y g X I u t
=
(24)
Where F and g are nonlinear functions, X is the system state variables, I is the inputs
variables, u is the control variables to be tuned and Y is the output.
3.2 Experimental validation
In this phase we used specified experimental data to identify the model parameters
used in the modeling process (models of representation: transfer function or neural
networks, models of knowledge,), and than we used dynamic experimental data to test
the model responses accuracy and its validation. This step is classic in any modeling
process.
Optimized Method for Real Time Nonlinear Control
169
3.3 Offline dynamic optimization
In this step we present the optimization technique of the control variables over dynamic
courses and we define the objective function to be used. The question that we should ask is
the following: In response to dynamic inputs I(t) which solicit the system over a certain
interval of time [0,T], what is the optimal continuous values of the control parameters u(t)
which minimize the cumulative production of the output Y(t). Therefore the objective
function to be minimized can be written using the following form:
0 0
( ) ( , , , )
i i
T T
a a
Min Y t dt Min g X I u t dt =
} }
(25)
The optimization problem has the following equalities and inequalities constraints:
Equalities constraints: ( , , , )
dX
F X I u t
dt
= (26)
Inequalities constraints:
min max
min max
X X X
u u u
< <
< <
(27)
Because the problem is nonlinear, there is no analytical solution; therefore we must
reformulate it into its discretized form as following:
Objective function:
1
( , , , )
i
N
i i i i i
a
i
Min g X I u t
=
(28)
Equality constraints:
1
( , , , ) ( , , , )
i i
i i i i i
X X dX
F X I u t F X I u t
dt t
+
= =
A
1
. ( , , , ) 0
i i i i i i
X X t F X I u t
+
A = (29)
Inequality constraints:
min max
min max
i
i
X X X
u u u
< <
< <
(30)
The inequality constraints are the domain definition of the systems state and control
variables; they are the lower and upper physical, mechanical or tolerance limits which
assure a good functioning performance of the system and prevent the system damage. In
our case, for example, the engine speed and the intake and exhaust pressure and
temperature must vary between a lower and upper limit to prevent engine system damage
or dysfunction.
3.4 Creation of the optimal database
The optimization problem explained previously necessitates a long computation time and
therefore it cannot be directly resolved in real time applications, in addition the inputs
evolution must be known beforehand which is not true in any real time applications.
Consequently we propose to resolve the problem off line for different inputs profiles that
Applications of Nonlinear Control
170
are very rich in information and variety and that cover a large area of possibility of the
systems domain and then to regroup the found solutions (inputs profiles and optimal
control variables) in a large database which will be exploited in the following step.
Therefore in the created database, we will find for every input vector I(t) an output vector
u(t) which is the optimal control variables that can be used to respond to the inputs
solicitations. In the next section this database will be used to create a dynamic controller
based on neural networks.
3.5 Online neural network control
Since a score of year, the neural networks are considered as a powerful mathematical tool to
perform nonlinear regression. Many engineers used them as a black box model to estimate
the system responses and they also used them in various fields of applications including
pattern recognition, forms recognition, objects classification, filters and control systems
(Rivals, 1995). We distinguish two main types of neural networks: feedforward or multi
layers networks used for steady state processes and feedback or recurrent networks used for
dynamic processes. We recognize to these networks the following fundamental
characteristics (Ouladssine, 2004): They are black box models with great capacity for
universal, flexible and parsimonious functions approximation.
We are interested in establishing a control technique by training a recurrent neural network
using the database created in the forth step of this methodology. The main advantage of this
approach is essentially the capacity of developing a nonlinear controller with a small
computation time which can be executed in real time applications.
Between the various neural networks architectures found in the literature, the multilayer
perceptrons are the most popular; they are particularly exploited in system modeling,
identification and control processes (Zweiri 2006). Many works show that the three layers
perceptrons with one hidden layer are universal function approximation (Li, 2005); they are
capable to approximate any nonlinear continuous function, defined from a finite multi
dimensions space into another, with an arbitrary fixed precision, while they require the
identification of a limited number of parameters comparing to the development of series of
fixed functions. In this way, they are parsimonious.
4. Application: Optimal air control in diesel engine
Many vehicles developers are especially interested in Diesel internal combustion engines
because of their high efficiencies reflecting low fuel consumption. Therefore, electronics and
common rail injection systems are largely developed and used in diesel engines along with
variable geometry turbocharger and exhaust gas recirculation in order to reduce the
pollution and protect the environment and the human health and to optimize the engine
performance and fuel consumption. The future engines must respect the more restricted
pollution legislations fixed in the European emissions standards (table 1). The particulate
matter that are mostly emitted under transient conditions due to air insufficiency are
expected to be reduced of a ratio 1:10 at 2010 (Euro 6) and the nitrogen oxides which are
caused by a smaller rate of the exhaust gas recirculation due to the insufficiency of fresh air
supplied to the engine by the compressor at low engine speed and fuel consumption
reduction and engine performance at high speed are also supposed to be reduced to half.
Optimized Method for Real Time Nonlinear Control
171
Heavy duty vehicle
Euro 1
1993
Euro 2
1996
Euro 3
2000
Euro 4
2005
Euro 5
2008
Euro 6
2010
Oxides nitrogen 9 7 5 3,5 2 1
Carbon monoxide 4,5 4 2,1 1,5 1,5 1,5
Hydrocarbons 1,23 1,1 0,66 0,46 0,46 0,46
Particulate Matter 0,4 0,15 0,1 0,02 0,02 0,002
Table 1. European standard of heavy duty vehicles in g/KW.h
Actually, modern diesel engines are controlled by look up tables which are the results of a
steady state optimization using experiments done on a test bench. Figure 1 shows a static
chart of the fresh air flow rate that is used to control the air management system. This chart,
as well as the entire look up tables used in the engine control, depends over two entries: the
crankshaft angular speed and the fuel mass flow rate (Arnold, 2007). The schematic
description of an open and closed loop control are shown in fig. 2 and 3. The inputs are the
pedals position X
p
and the engine angular speed w. The outputs are the actuators of the
turbine variable geometry GV and the opening position of the exhaust gas recirculation
EGR. The indication ref designates a reference value and the indication corr is its corrected
value. P and m
C
are respectively the predicted or measured intake pressure and the air mass
flow rate entering the intake manifold.
Fig. 1. Static chart of the fresh air flow rate used in the engine control schemes.
Fig. 2. Open loop control
Engine
Optimal
static
maps
e
GV
ref
EGR
ref
Dynamic
corrector
Fuel
Opacity
GV
cor
EGR
cor
Applications of Nonlinear Control
172
Fig. 3. Closed loop control
In the open loop control, the classic control of a diesel engine (Hafner, 2000) is done
according to the diagram in fig. 2, the optimal values of the actuators are updated by
memorized static maps. Then a predictive corrector (Hafner, 2001) is generally used in order
to compensate the engine dynamic effects.
In the closed loop control (fig. 3), the engine is controlled by error signals which are the
difference between the predicted or measured air mass flow rate and the intake pressure,
and their reference values. The controller uses memorized maps as reference, based on
engine steady state optimization (Hafner, 2001; Bai, 2002). The influence of the dynamic
behavior is integrated by several types of controller (PI, robust control with variable
parameters, ) (Jung, 2003).
Our work proposes practical solutions to overcome and outperform the control insufficiency
using static maps. The advantage of this approach is to be able to propose dynamic maps
capable of predicting, on line, the inair cylinders filling. Therefore the optimal static maps
in fig. 1 and 2 can be replaced by optimal dynamic ones.
We suggest a mathematical optimization process based on the mean value engine model to
minimize the total pollutants production and emissions over dynamic courses without
deteriorating the engine performance. We used the opacity as a pollution criterion, this choice
was strictly limited due to the available data, but the process is universal and it can be applied
individually to each pollutant which has physical model or to the all assembled together.
This optimizations procedure is difficult to be applied directly in on line engines
applications, due to the computation difficulties which are time consuming. Consequently, it
will be used to build up a large database in order to train a neural model which will be used
instead. Neural networks are very efficient in learning the nonlinear relations in complex
multivariables systems; they are accurate, easy to train and suitable for real time applications.
All the simulations results and figures presented in this section were computed using
Matlab development environment and toolboxes. The following section is divided to four
subsections as follows: I Engine dynamic modeling, II Simulation and validation of the
engines model, III Optimization over dynamic trajectories, IV Creation of Neural network
for on line controller.
4.1 Engine dynamic modeling
Diesel engines can be modeled in two different ways: The models of knowledge quasistatic
(Winterbonne, 1984), drainingreplenishment (Kao, 1995), semi mixed (OuenouGamo, 2001;
Younes, 1993), bond graph (Hassenfolder, 1993), and the models of representation by
transfer functions (Younes, 1993), neural networks (Ouladssine, 2004).
c
P.I
Engine
Optimal
static
maps
e
m'C, ref
Pref
mC
P
Opacity Fuel GV
EGR
Optimized Method for Real Time Nonlinear Control
173
Seen our optimization objective, the model of knowledge will be adopted in this work. The
semimixed model is the simplest analytic approach to be used in an optimization process.
The Diesel engine described here is equipped with a variable geometry turbocharger and
water cooled heat exchanger to cool the hot air exiting the compressor, but it doesnt have
an exhaust gas recirculation system that is mainly used to reduce the NOx emissions.
Consequently the engine is divided to three main blocks: A. the intake air manifold, B. the
engine block, C. the opacity (Omran, 2008a).
4.1.1 Intake air manifold
Considering air as an ideal gas, the state equation and the mass conservation principle gives [4]:
( )
a
a a c a0
dP
V r.T . m m
dt
= (31)
c
m is the compressor air mass flow rate,
0 a
m is the air mass flow rate entering the engine,
P
a
, V
a
and T
a
are respectively the pressure, the volume and the temperature of the air in the
intake manifold and r is the mass constant of the air.
0 a
m is given by:
0 0, a V a th
m m q = (32)
0, a th
m is the theoretical air mass flow rate capable of filling the entire cylinders volume at
the intake conditions of pressure and temperature:
a
a0,th
a
Vcyl..P
m
4 r T
=
(33)
V
cyl
is the displacement, is the crankshaft angular speed, and
v
is the inair filling
efficiency given by:
2
v 0 1 2
= + +
(34)
Where
i
are constants identified from experimental data. The intake temperature T
a
is
expressed by:
( )
a ech c ech water
T 1 T T = +
(35)
T
c
is the temperature of the air at the compressors exit. T
water
is the temperature of the
cooling water supposed constant.
ech
q is the efficiency of the heat exchanger supposed
constant. The temperature T
c
is expressed by:
1
a
c 0
0 c
P 1
T T 1 1
P
   
 
 
= +

 
\ .  
\ . \ .
(36)
4.1.2 Engine block
The principle of the conservation of energy applied to the crankshaft gives:
Applications of Nonlinear Control
174
( )
2
e r
d 1
J
dt 2
 
= P P

\ .
(37)
J() is the moment of inertia of the engine, it is a periodic function of the crankshaft angle due
to the repeated motion of its pistons and connecting rods, but for simplicity, in this paper, the
inertia is considered constant. P
e
is the effective power produced by the combustion process:
e e f ci
.m .P P =
(38)
f
m is the fuel flow rate, P
ci
is the lower calorific power of fuel and
e
is the effective
efficiency of the engine modeled by [5]:
2
1 2 3 4
e
2 2 2 2
5 6 7
c c c c w
c w c w c w
 
+ + + +
 =

+ +
\ .
(39)
c
i
are constants, and is the coefficient of air excess:
a0
f
m
m
=
(40)
P
r
is the resistant power:
r r
C P =
(41)
Cr is the resistant torque. Fig. 4 represents a comparison between the effective efficiency
model and the experimental data measured on a test bench. The model results are in good
agreement with experimental data.
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Air Excess
E
f
f
e
c
t
i
v
e
E
f
f
e
c
i
e
n
c
y
Model at 800 RPM
Model at 1200 RPM
Model at 1600 RPM
Model at 2000 RPM
Exp. Data at 800 RPM
Exp. Data at 1200 RPM
Exp. Data at 1600 RPM
Exp. Data at 2000 RPM
Fig. 4. Comparison between the effective efficiency model results and the experimental data
at different crankshaft angular speed.
Optimized Method for Real Time Nonlinear Control
175
4.1.3 Diesel emissions model
The pollutants that characterize the Diesel engines are mainly the oxides of nitrogen and the
particulate matters. In our work, we are especially interested in the emitted quality of
smokes which is expressed by the measure of opacity (Fig. 5) (OuenouGamo, 2001):
4 5 6 3 2
m w m m w m m
1 a f
Opacity m w m m
+ +
=
(42)
m
i
are constants identified from the experimental data measured over a test bench.
15
20
25
30
35
40
45
50
55
60 750
1000
1500
2000
2500
0
20
40
60
80
100
Opacity
[%]
Air/Fuel Ratio
Crankshaft Angular
Speed [rpm]
Fig. 5. Graphical representation of the opacity computed using (32) and a constant fuel flow
rate equal to 6 g/s.
4.1.4 System complete model
Reassembling the different blocks equations leads to a complete model describing the
functioning and performance of a variable geometry turbocharged Diesel engine. The model
is characterized by two states variables (P
a
, w), two inputs (
f
m , C
r
) and the following two
differential equations representing the dynamic processes:
( )
a
a a c ao
2
e f ci r
dP
V r T m m
dt
d 1
J m P C w
dt 2
q
 
=

\ .
(43)
4.2 Model validation
The test bench, conceived and used for the experimental study, involves: a 6 cylinders
turbocharged Diesel engine and a brake controlled by the current of Foucault. Engines
characteristics are reported in table 2.
Applications of Nonlinear Control
176
Stroke [mm] 145
Displacement [cm3] 9839.5
Volumetric ratio 17/1
Bore [mm] 120
Maximum Power [KW]
at crankshaft angular speed [rpm]
260
2400
Maximum torque [daN.m]
at crankshaft angular speed [rpm]
158
1200
Relative pressure of overfeeding [bar] 2
Table 2. Engine Characteristics
Different systems are used to collect and analyze the experimental data in transient phase
and in real time functioning:  Devices for calculating means and instantaneous measures, 
a HC analyzer by flame ionization,  a Bosch smoke detector and  an acquisition device for
signal sampling. The use of these devices improves significantly the quality of the static
measures by integration over a high number of points.
Fig. 6 and 7 show a comparison between two simulations results of the engine complete
model and the experimental data. The inputs of the model are the fuel mass flow rate
and the resistant torque profiles. The output variables are: the pressure of the intake
manifold P
a
the crankshaft angular speed and the opacity characterizing the engine
pollution. The differential equations described in section 4.1.4 are computed simultaneously
using the RungeKutta method. The simulations are in good agreement with the
experimental data.
0 50 100
0
5
10
t [s]
F
u
e
l
[
g
/
s
]
Inputs data
0 50 100
0
500
1000
1500
t [s]
R
e
s
i
s
t
a
n
t
t
o
r
q
u
e
[
N
.
m
]
0 50 100
1000
1500
2000
t [s]
E
n
g
i
n
e
s
p
e
e
d
[
r
p
m
]
Simulation results
Experimental data
0 50 100
1
1.5
2
2.5
t [s]
I
n
t
a
k
e
p
r
e
s
s
u
r
e
p
a
[
b
a
r
]
0 50 100
0
10
20
30
t [s]
O
p
a
c
i
t
y
[
%
]
Fig. 6. Simulation 1: Comparison between the complete engine model and the experimental
data measured on the test bench.
Optimized Method for Real Time Nonlinear Control
177
0 100 200 300
0
5
10
15
t [s]
F
u
e
l
[
g
/
s
]
Inputs data
0 100 200 300
0
500
1000
1500
t [s]
R
e
s
i
s
t
a
n
t
t
o
r
q
u
e
[
N
.
m
]
0 100 200 300
1500
2000
2500
t [s]
E
n
g
i
n
e
s
p
e
e
d
[
r
p
m
]
Simulation results
Experimental data
0 100 200 300
1
1.5
2
2.5
t [s]
I
n
t
a
k
e
p
r
e
s
s
u
r
e
p
a
[
b
a
r
]
0 100 200 300
6
8
10
12
t [s]
O
p
a
c
i
t
y
[
%
]
Fig. 7. Simulation 2: Comparison between the complete engine model and the experimental
data measured on the test bench.
4.3 Optimization process
4.3.1 Problem description
When conceiving an engine, engines developers have always to confront and solve the
contradictory tasks of producing maximum power (or minimum fuel consumption) while
respecting several pollutions constraints (European emissions standard). We are only
interesting in reducing the pollutants emissions at the engine level, by applying the optimal
inair cylinders filling. Consequently, the problem can now be defined; it consists in the
following objective multicriteria function:
Maximize "Power"
Minimize "Pollutants"
(44)
This multiobjective optimization problem can be replaced by a single, non dimensional,
mathematical function regrouping the two previous criteria:
max ,max
i
i i
Poll
f dt dt
Poll
P
= +
`
P
)
} }
(45)
P is the engine effective power, Poll
i
is a type of pollutant, and the indication max
characterizes the maximum value that a variable can reach. The integral represents the heap
of the pollutants and power over a given dynamic trajectory. This trajectory can be, as an
example, a part of the New European Driving Cycle (NEDC).
In this chapter we will only use the opacity as an indication of pollution seen the simplicity
of the model and the priority given to the presentation of the method, but we should note
that the optimization process is universal and it can involve as many pollutions criteria as
we want. The function "objective" becomes:
Applications of Nonlinear Control
178
max max
Op
f dt dt
Op
P
= +
P
} }
(46)
4.3.2 Formulation of the problem
The problem consists therefore in minimizing the following function "objective" over a
definite working interval [0, t]:
4
max
1
max
5 6 3 2
ci
e f
m w m m w m m
a f
P
m dt
P
f
m
w m m dt
Op
q
+ +
=
`
+
)
}
}
(47)
Under the equalities constraints representing the differential equations of the engine block
and the intake manifold:
( )
a
a a c ao
2
e f ci r
dP
V r T m m
dt
d 1
J m P C w
dt 2
q
 
=

\ .
(48)
And the inequalities constraints derived from the physical and mechanical limits of the air
excess ratio, the intake pressure and the crankshaft angular speed:
 
 
15 80
.5
4 4
a
9 .10 P 30.10 Pa
83 230 rd / s
s s
s s
s s
(49)
is given by:
( )
2
0 1 2 a
f
Ncyl.Vcyl..P
4 m
+ +
=
(50)
The variables of the optimizations problem are w, P
a
and m
c
, and the inputs are C
r
and m
f
.
We should note that we intentionally eliminate the exhaust and turbocompressor blocks
from the equalities constraints because we are interesting in obtaining the optimal inair
cylinders filling m
c
without being limited to any equipments such as the variable geometry
turbocompressor early described. In other words, we can consider that we have replaced
the turbocompressor by a special instrument that can deliver to the intake manifold any
value of the air mass flow rate that we choose and at any time. Later, in the conclusion, the
devices that can provide these optimal values are briefly discussed. Also we should note
that we will use the complete engine model of the existing turbocharged diesel engine as a
comparison tool, to prove the validity of our proposed optimal control and the gain in the
opacity reduction.
Optimized Method for Real Time Nonlinear Control
179
4.3.3 Problem discretization
There is no analytic solution to the problem previously formulated; therefore there is a
necessity to reformulate it in its discretized form. The integrals in the function "objective"
become a simple sum of the functions computed at different instant t
i
:
1
N
i 1 2 N
i
f f f f f
=
= = + +
(51)
N is the number of the discretized points, it is the size of the unknown vectors e
,
a
P
and
c
m
.
h is the step of discretization. Using the Taylor development truncated at the first
differential order, the equalities constraints become:
( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( )
( )
a a i 1 a i c i ao i
a
2 2
e i 1 i i r i
h
P P r.T . m m 0
V
2 h
0
J
+
+
P P =
(52)
And the inequalities constraints:
( )
( )
 
( )
 
15 80
.5
i
4 4
a i
i
9 .10 P 30.10 Pa
83 230 rd / s
s s
s s
s s
(53)
4.3.4 Solution of the optimization problem
The optimization problem under equality and inequality constraints can be described using
the following mathematical form:
( ) { }
( )
( )
( )
1 2
, ,...
int
0 i 1,...,m
0 i 1,...,p
n
i
i
Min f X
X x x x
Under Constra s
h X
g X
= =
s =
(54)
Where f(X) is the objective function, h(X) the equality constraints and g(X) the inequality
constraints. The easiest way to resolve this problem is to reduce it to a problem without
constraints by creating a global objective function (X, h(X), g(X)) which regroups the
original objective function and the equality and inequality constraints (Minoux, 1983).
Therefore we will use a global objective function that regroups: The function objective, the
equalities constraints with Lagrange multipliers, and the inequalities constraints with a
penalty function. The final objective function becomes (Minoux, 1983):
( ) ,
p
m
2
i i i
i 1 i 1
L X f(X) * h (X) r. [g (X)]
= =
= + +
(55)
Applications of Nonlinear Control
180
0
k
r r = , k is the number of iteration that must tend toward the infinity, and r
0
= 3. The
problem will have m additional unknown variables (Lagranges multipliers
i
) to be
determined along with the engines variables. The algorithm of the minimization process
adopted here is the BroydenFletcherGoldfarbShanno B.F.G.S. that sums up as follows:
1. To start by an initial solution
( ) 0
X .
2. To estimate the solution at the k iteration by: ( )
1 k k k k k
X X D f X o
+
= V , with X is a
vector regrouping the optimization variables,
k
is a relaxation factor, ( ) { }
k k
D f X V
represents the decreasing direction of the function,
1
k
D
is an approximation of the
Hessian matrix.
3. To verify if the gradients module of the objective function at the new vector X is under
a certain desired value ( 10
2
). If it is true then this solution is the optimal solution, end
of search. Otherwise increment k and return to the stage 2.
4.3.5 Results and discussion
We applied the optimization process explained in the previous section to two different
profiles of inputs variables (fuel mass flow rate and resistant torque). The time step of
discretization h is equal to 0.01s and the time interval is equal to 3 sec, each problem has 900
unknown variables { } , and
a c
w P m with 598 equalities constraints and 1800 inequalities
constraints. Fig. 8 and 9 show a comparison between the results of the optimization process
and the simulations results of the engines complete system model for the same input
values and at fixed position of the turbine variable geometry (completely open, GV = 0). The
optimizations results show that we need significantly more air mass flow rate entering the
intake manifold and higher intake pressure to reduce the opacity, while the real
turbocharged diesel engine is not capable of fulfilling these tasks.
0 1 2 3
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
t [s]
m
'
c
[
K
g
/
s
]
Engine Complete Model's simulation
Optimization process' results
0 1 2 3
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t [s]
i
n
t
a
k
e
P
r
e
s
s
u
r
e
P
a
[
B
a
r
]
Engine Complete Model's simulation
Optimization process' results
Fig. 8. Comparison between the air mass flow rate and the intake pressure calculated using
the optimization procedure and the ones simulated using the engine complete model for a
variable fuel flow rate and a constant resistant torque equal to 1000 N.m.
Optimized Method for Real Time Nonlinear Control
181
0 1 2 3
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
t [s]
m
'
C
[
K
g
/
s
]
Engine Complete Model's simulation
Optimization process' results
0 1 2 3
1
1.5
2
2.5
t [s]
I
n
t
a
k
e
P
r
e
s
s
u
r
e
P
a
[
B
a
r
]
Engine Complete Model's simulation
Optimization process' results
Fig. 9. Comparison between the air mass flow rate and the intake pressure calculated using
the optimization procedure and the ones simulated using the engines complete model for a
variable fuel flow rate and a variable resistant torque.
Fig. 10 and 11 show a comparison between the simulated opacity derived from the
optimization process and the one derived from the engines complete system model for the
same inputs used in fig. 8 and 9. The enormous gain in the opacity reduction proves the
validity of the suggested optimization procedure.
0 0.5 1 1.5 2 2.5 3
0
10
20
30
40
50
60
t [s]
O
p
a
c
i
t
y
[
%
]
Engine Complete Model's simulation
Optimization process' simulation
Fig. 10. Opacity reduction using the optimal values of the air mass flow rate and the intake
pressure. Blue: simulation without optimization, red: Simulation with optimization.
Applications of Nonlinear Control
182
0 0.5 1 1.5 2 2.5 3
0
5
10
15
20
25
30
35
40
t [s]
O
p
a
c
i
t
y
[
%
]
Engine Complete Model's simulation
Optimization process' simulation
Fig. 11. Opacity reduction using the optimal values of the air mass flow rate and the intake
pressure. Blue: simulation without optimization, red: Simulation with optimization.
4.4 Neural network controller
Optimization previously done "off line", would be directly unexploited "online" by a
controlling processor seen the enormous computation time that is necessary to resolve the
optimization problem. In order to integrate the results of this optimizations procedure in a
closed loop controller (ref fig. 3), and to be able to use it in real time engine applications, we
suggest to use a black box model based on neurons. Neural network is a powerful tool
capable of simulating the engines optimal control variables with good precision and almost
instantly.
The neural network inputs are the fuel mass flow rate and the resistant torque, and its
output variables are the optimal values of the air mass flow rate and the intake pressure.
However in real time engine applications, the injected fuel flow rate is measurable, while the
resistant torque is not. Consequently, we suggest substituting this variable by the crankshaft
angular speed which can be easily measured and which is widely used in passenger cars
controlling systems.
Firstly, we need to create a large database which will be used to train the neural model, and
which covers all the functioning area of the engine in order to have a good precision and a
highly engine performance. The database is created using the optimization process as
explained in subsection 4.3.
Then we have to judicially choose the number of the inputs time sequence to be used, in
order to capture the inputs dynamic effects and accurately predict the output variables.
With intensive simulations and by trial and error, we find out that a neural network with
inputs the fuel mass flow rate and the crankshaft angular speed at instant (i), (i1) and (i2) is
capable of precisely predicting the optimal values of the air mass flow rate and the intake
Optimized Method for Real Time Nonlinear Control
183
pressure at current instant (i). Fig. 12 describes the neural network. The network is built
using one hidden layer and one output layer, the activation functions of the hidden layer are
sigmoid; the ones at the output layer are linear.
Fig. 12. The structural design of the neural network adopted in this paper for predicting the
optimal control of the inair filling and the intake pressure in real time applications.
The number of neurons in the hidden layer is determined by referring to the errors
percentage of the points which are under a certain reference value wisely chosen; the
errors percentage (table 3) are the results of the difference between the outputs of the
network after the training process is completed, and the desired values used in the training
database.
Table 3 shows the results of the neural networks with different number of neurons in their
hidden layer, these networks are trained with the same database until a mean relative error
equal to 10
8
is reached or maximum training time is consumed. The values in the table
represent the percentage of the neural network results respecting the specified error
percentage computed with respect to the reference values.
Number of neurons of the
hidden layer
Error percentage
Relative
error
< 1 % < 5 %
< 10
%
110 57.71 88.85 96.71 3.6 10
5
120 98.428 100 100 10
8
130 98.734 100 100 10
8
140 99 100 100 10
8
Table 3. Results of four neural networks trained using different neurons number in their
hidden layer and the same database.
The neural network adopted in this paper includes one hidden layer with 140 neurons and
one output layer with 2 neurons. Fig. 13 and 14 show a comparison between the air mass
flow rate and the intake pressure calculated using the theoretical optimization procedure,
and the ones computed using the neural network. The results are almost identical; the mean
relative error is 10
6
.
) 1 (
.
i
f
m
) (
.
i
f
m
) (i w
) (
.
i mc
) 2 (
.
i
f
m
) 1 ( i w
) 2 ( i w
.
.
.
.
.
) (i p
c
Applications of Nonlinear Control
184
0 1 2 3
0.28
0.3
0.32
0.34
0.36
0.38
0.4
0.42
t [s]
m
'
C
[
K
g
/
s
]
Optimization Process
Neural Model
0 1 2 3
2.32
2.34
2.36
2.38
2.4
2.42
2.44
2.46
t [s]
I
n
t
a
k
e
P
r
e
s
s
u
r
e
[
B
a
r
]
Optimization Process
Neural Model
0 1 2 3
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t [s]
m
'
C
[
K
g
/
s
]
Optimization Process
Neural Model
0 1 2 3
1.7
1.8
1.9
2
2.1
2.2
2.3
2.4
2.5
t [s]
I
n
t
a
k
e
P
r
e
s
s
u
r
e
[
B
a
r
]
Optimization Process
Neural Model
Fig. 13. & 14. Comparison between the neural network outputs and the optimal values of the
air mass flow rate and the intake pressure.
5. General conclusions
We successfully developed and validated a mean value physical model that describes the
gas states evolution and the opacity of a diesel engine with a variable geometry
turbocharger. Then we proposed a dynamic control based on the optimal inair cylinders
filling in order to minimize the pollutants emissions while enhancing the engine
performance. The optimization process is described in detail and the simulation results (fig.
811) prove to be very promising. In addition, the control principle as described here with
the opacity criterion can be easily applied to other pollutants which have available physical
model. This will be the object of future publications.
Optimized Method for Real Time Nonlinear Control
185
Also, in order to overcome on line computation difficulties, a real time dynamic control based
on the neural network is suggested; therefore the optimal static maps of the fig. 2 can be
successfully replaced by dynamic maps simulated in real time engine functioning (fig. 15).
Fig. 15. Proposed control in closed loop
Finally, we should note that, in this chapter, while we did find, in theory, the optimal air mass
flow rate and intake pressure necessary to minimize the opacity, but we didnt discuss the
mechanical equipments required to provide the optimal intake pressure and intake air flow
rate in real time engine applications. The practical implementation of the dynamic control is an
important question to be studied thereafter. The use of a turbocompressor with variable
geometry and/or with WasteGate, and/or electric compressor is to be considered.
6. References
Arnold, J.F. (2007). Proposition dune stratgie de commande base de logique floue pour la commande
du circuit dair dun moteur Diesel. PHD thesis, Rouen University, (March 2007), France.
Bai, L. & Yang, M. (2002). Coordinated control of EGR and VNT in turbocharged Diesel engine
based on intake air mass observer, SAE Technical paper 2002011292, (March 2002).
Bartoloni, G. (1989). Chattering Phenomena in Discontinuous Control Systems. International
Journal on Systems Sciences, Vol. 20, Issue 12, (February 1989), ISSN 00207721.
Bellman, R. (1975). Dynamic programming. Princeton University Press, (1957), Princeton, NJ.
Doyle, J.C. (1979). Robustness of multiloop linear feedback systems Proceedings of the 1978
IEEE Conference on Decision and Control, pages 1218, Orlando, December 1979.
Dreyfus, S.E. (1962). Variational Problems with Inequality Constraints. Journal of
Mathematical Analysis and Applications, Vol. 4, Issue 2, (July 1962), ISSN 0022247X.
Friedland, B. (1996). Advanced Control System Design. PrenticeHall, Englewood Cliffs, ISBN
9780130140104, NJ, 1996.
Hafner, M. (2000). A NeuroFuzzy Based Method for the Design of Combustion Engine
Dynamometer Experiments. SAE Technical Paper 2000011262, (March 2000).
Hafner, M. (2001). Model based determination of dynamic engine control function
parameters. SAE Technical Paper 2001011981, (May 2001).
Hahn, W. (1967). Stability of Motion, SpringerVerlag, ISBN 9783540038290, New York, 1967.
Hassenfolder, M. & Gissinger, G.L. (1993). Graphical eider for modelling with bound graphs
in processes . ICBGM93, pp. 188192, Californie, January 1993.
Jung, M. (2003). Mean value modelling and robust control of the airpath of a turbocharged diesel
engine. Thesis for doctor of philosophy, University of Cambridge, 2003.
Kao, M. & Moskwa, J.J. (1995). Turbocharger Diesel engine modelling for non linear engine
control and state estimation. Trans ASME, Journal of Dynamic Systems Measurement
and Control, Vol. 117, Issue 1, pp. 2030, (March 1995), ISSN 00220434.
c
P.I
Engine
Neural
Network
e
m'
C, ref
P
ref
m
C
P
Opacity Xp GV
EGR
Applications of Nonlinear Control
186
Li, D.; Lu, D.; Kong X. & Wu G. (2005). Implicit curves and surfaces based on BP neural
network. Journal of Information & Computational Science, Vol. 2, No 2, pp. 259271,
(2005), ISSN 17467659.
Minoux, M. (1983). Programmation Mathmatique, Thorie et Algorithmes. tome 1 & 2, editions
dunod, ISBN 9782743010003, Paris 1983.
Omran, R.; Younes, R. & Champoussin, J.C. (2008a). Optimization of the InAir Cylinders
Filling for Emissions Reduction in Diesel engines, SAE Technical Paper 2008011732,
(June 2008).
Omran, R. ; Younes, R. & Champoussin, J.C. (2008b). Neural Networks for Real Time non
linear Control of a Variable Geometry Turbocharged Diesel Engine, International
Journal of Robust and non Linear Control, Vol. 18, Issue 12, pp. 12091229, (August
2008), ISSN 10991239.
OuenouGamo, S. (2001). Modlisation dun moteur Diesel suraliment, PHD Thesis, Picardie
Jules Vernes University, (2001), France.
Ouladssine, M.; Blosh, G. & Dovifazz X. (2004). Neural Modeling and Control of a Diesel
Engine with Pollution Constraints. Journal of Intelligent and Robotic Systems; Theory
and Application, Vol. 41, Issue 23, (January 2005), ISSN 09210296.
Outbib, R. & Vivalda, J.C. (1999). A note on Feedback stabilization of smooth nonlinear
systems. IEEE Transactions on Automatic Control. Vol. 44, No 1, pp. 200203, (August
1999), ISSN 00189286.
Outbib, R. & Richard, E. (2000). State Feedback Stabilization of an Electropneumatic System.
ASME Journal of Dynamic Systems Measurement and Control. Vol. 122, pp 410415,
(September 2000), ISSN 00220434.
Outbib, R. & Zasadzinski, M. (2009). Sliding Modes Control, In: Control Methods for Electrical
Machines, Ren Husson Editor, Ch. 6, pp. 169204, 2009. Wiley, ISBN 978184821
0936, USA.
Pontryagin, L.S.; Boltyanskii, V.; Gamkrelidze, R.V & Mishchenko, E.F. (1962). The
Mathematical Theory of Optimal Processes. 1962, John Wiley & Sons, USA
Rivals, I. (1995). Modlisation et commande de processus par rseaux de neurones ; application au
pilotage dun vhicule autonome, PHD Thesis, Paris 6 University, (1995), France.
SiraRamirez, H. (1987). Differential Geometric Methods in VariableStructure Control.
International Journal of Control, Vol. 48, Issue 2, (March 1987), pp. 13591390, ISSN
00207179.
Slotine, JJ.E. (1984). Sliding Controller Design for Nonlinear System. International Journal of
Control, Vol. 40, Issue 2, pp. 421434, (March 1984), ISSN 00207179.
Utkin, V.I. (1992). Sliding Modes in Control and Optimization, SpringerVerlag, ISBN 978
0387535166, Berlin 1992.
Winterbonne, D.E. & Horlock, J.H. (1984). The thermodynamics and gas dynamics internal
combustion engines. Oxford Science Publication, ISBN 9780198562122, London 1984.
Younes, R. (1993). Elaboration dun modle de connaissance du moteur diesel avec
turbocompresseur gomtrie variable en vue de loptimisation de ses missions. PHD
Thesis, Ecole Centrale de Lyon, (November 1993), France.
Zames, G. (1981). Feedback and optimal sensitivity: Model reference transformations,
multiplicative seminorms, and approximative inverse IEEE Transcations on
Automatic Control, Vol. 26, Issue 2, pp. 301320, (June 1981), ISSN 00189286.
Zweiri Y.H. (2006). Diesel engine indicated torque estimation based on artificial neural
networks. International Journal of Intelligent Technology, Vol. 1, No 1, pp. 233239,
(July 2006), ISSN 13056417.
0
Nonlinear Phenomena and Stability Analysis for
Discrete Control Systems
Yoshifumi Okuyama
Tottori University, Emeritus
Japan
1. Introduction
Almost all feedback control systems are realized using discretized (discretetime and
discretevalue, i.e., digital) signals. However, the analysis and design method of
discretized/quantized (nonlinear) control systems has not been established (Desoer et al.,
1975; Elia et al., 2001; Harris et al., 1983; Kalman, 1956; Katz, 1981). This article analyzes
the nonlinear phenomena and stability of discretized control systems in a frequency domain
1
(Okuyama, 2006; 2007; 2008). In these studies, it is assumed that the discretization is
executed on the input and output sides of a nonlinear element at equal spaces, and the
sampling period is chosen of such a size suitable for the discretization in the space. Based
on the premise, the discretized (pointtopoint) nonlinear characteristic is examined from two
viewpoints, i.e., global and local. By partitioning the discretized nonlinear characteristic into
two sections and by dening a sectorial area over a specied threshold, the concept of the
robust stability condition for nonlinear discretetime systems is applied to the discretized
(hereafter, simply wrriten as discrete) nonlinear control system in question. As a result, the
nonlinear phenomena of discrete control systems are claried, and the stability of discrete
nonlinear feedback systems is elucidated.
N(e
)
G(s)
T1
T1
e
e
r
r
g
g E E
c
' '
T
r
h
h
S
1
S
2
v
u
y
d
+
+
+
Fig. 1. Nonlinear sampleddata control system.
2. Discrete nonlinear control system
The discrete nonlinear control system to be considered here is represented by a sampleddata
control system with two samplers, S
1
, S
2
and the continuous nonlinear characteristic N() as
1
In the time domain analysis (e.g., Lyapunov function method), it is difcult to nd a Lyapunov function
for the discretized (severe nonlinear characteristic) feedback system. The frequency domain analysis
will be important in cases where physical systems with uncertainty in the systemorder are considered.
11
2 Nonlinear Control
N(e
) T
1
T
2
G(z)
g
g
E E E E
c
' '
T
r e e
y u
v
d
+
+
N
d
(e)
Fig. 2. Discrete nonlinear control system.
(a) (b)
Fig. 3. Discretized nonlinear characteristics.
shown in Fig. 1. Here, T1 denotes the discretization and zeroorderhold, which are usually
performed in A/D(D/A) conversion, and G(s) is the transfer function of the linear controlled
system. It is assumed that the two samplers with a sampling period h operate synchronously.
The sampleddata control system can be equivalently transformed into a discrete control
systemas shown in Fig. 2. Here, G(z) is the ztransformof G(s) together with zeroorderhold,
and T
1
and T
2
are the discretizing units on the input and output sides of the nonlinear
element, respectively. The relationship between e and v
= N
d
(e) in the gure becomes a
stepwise nonlinear characteristic on integer grid coordinates as shown in Fig. 3 (a). Here, a
rounddown discretization, which is usually executed on a computer, is applied. Therefore,
the relationship between e
and u
, u
, indicates a
188 Applications of Nonlinear Control
Nonlinear Phenomena and Stability Analysis for Discrete Control Systems 3
discrete value that can be assigned to an integer number, e.g.,
e
, 3, 2, , 0, , 2, 3, ,
u
, 3, 2, , 0, , 2, 3, ,
where is the resolution of each variable. Here, it is assumed that the input and output signals
of the nonlinear characteristic have the same resolution in the discretization. In the gure,
e
and u
(k) and u
(k), u
(k) Z
Z
def
= , 3, 2, 1, 0, 1, 2, 3, .
3. Equivalent discretetime system
In this study, the stepwise and pointtopoint nonlinear characteristic is partitioned into the
following two sections:
N
d
(e) = K(e + (e)), 0 < K < ,
[(e)[ < , (1)
for [e[ < , and
N
d
(e) = K(e + n(e)), 0 < K < ,
[n(e)[ [e[, 0 < 1, (2)
for [e[ , where (e) and n(e) are nonlinear terms relative to nominal linear gain K. Equation
(1) represents a bounded nonlinearity which exists in a nite region. On the other hand, (2)
represents a sectorial nonlinearity of which the equivalent linear gain exists in a limited range.
Therefore, when we consider the robust stability in a global sense, it is sufcient to consider
the nonlinear term n(e). Here, is a threshold of the input signal e. As a matter of course, (1)
and (2) must be satised with respect to the discretized value e = e
because e
e.
Based on the above consideration, the following newsequences e
m
(k) and w
m
(k) are dened:
e
m
(k) = e
m
(k) + q
e
(k)
h
, (3)
w
m
(k) = w
m
(k) q
e
(k)
h
. (4)
where q is a nonnegative number, e
m
(k) and w
m
(k) are neutral points of sequences e
(k) and
w
(k),
e
m
(k) =
e
(k) + e
(k 1)
2
, (5)
w
m
(k) =
w
(k) + w
(k 1)
2
, (6)
189 Nonlinear Phenomena and Stability Analysis for Discrete Control Systems
4 Nonlinear Control
1 + q n
()
q
` E E E f E
T
E
e e
w
n(e)
+
+
Fig. 4. Nonlinear subsystem.
n
()
F(, q, z)
g
g
E E
c
' '
T
r
/
e
u
/
y
/
d
/
+
+
+
Fig. 5. Equivalent feedback system.
and e
(k),
e
(k) = e
(k) e
(k 1). (7)
The relationship between equations (3) and (4) in regard to the continuous values is shown by
the block diagram in Fig. 4. In this gure, is dened as
(z) :=
2
h
1 z
1
1 + z
1
. (8)
Equation (8) corresponds to the bilinear transformation between z and . Thus, the loop
transfer function from w
to e
, u
/
and y
/
written
in Fig. 5 indicate the ztransformed ones.
In this study, the following assumption is provided on the basis of the relatively fast sampling
and the slow response of the controlled system.
[Assumption] The absolute value of the backward difference of sequence e(k) is not more
than , i.e.,
[e(k)[ = [e(k) e(k 1)[ . (10)
If the condition (10) is satised, e
(k)[ = [e
(k) e
(k 1)[ = or 0.
This assumption will be satised in the following examples.
190 Applications of Nonlinear Control
Nonlinear Phenomena and Stability Analysis for Discrete Control Systems 5
4. Norm conditions
In this section, some lemmas for norm conditions are presented. Here, in regard to (2), the
following new nonlinear function is dened.
f (e) := n(e) + e. (11)
When considering the discretized output of the nonlinear term, w
= n(e
), the following
expression can be given:
f (e
(k)) = w
(k) + e
(k). (12)
From inequality (2), it can be seen that the function (12) belongs to the rst and third
quadrants. Considering the equivalent linear characteristic, the following inequality can be
dened:
0 (k) :=
f (e
(k))
e
(k)
2. (13)
When this type of nonlinearity (k) is used, inequality (2) can be expressed as
w
(k) = n(e
(k)) = ((k) )e
(k). (14)
For the neutral points of e
(k) and w
(k)) + f (e
(k 1))) = w
m
(k) + e
m
(k). (15)
Moreover, equation (14) is rewritten as
w
m
(k) = ((k) )e
m
(k).
Since [e
m
(k)[ [e
m
(k)[, the following inequality is satised when a rounddown discretization
is executed:
[w
m
(k)[ [e
m
(k)[ [e
m
(k)[. (16)
Based on the above premise, the following norm inequalities are examined (Okuyama et al.,
1999; Okuyama, 2006).
[Lemma1] The following inequality holds for a positive integer p:
w
m
(k)
2,p
e
m
(k)
2,p
e
m
(k)
2,p
. (17)
Here,  
2,p
denotes the Euclidean norm, which can be dened by
x(k)
2,p
:=
_
p
k=1
x
2
(k)
_
1/2
.
(Proof) The proof is clear from inequality (16).
[Lemma2] If the following inequality is satised in regard to the inner product of the neutral
points of (12) and the backward difference (7):
w
m
(k) + e
m
(k), e
(k)
p
0, (18)
191 Nonlinear Phenomena and Stability Analysis for Discrete Control Systems
6 Nonlinear Control
the following inequality can be obtained:
w
m
(k)
2,p
e
m
(k)
2,p
(19)
for any q 0. Here, ,
p
denotes the inner product, which can be dened as
x
1
(k), x
2
(k)
p
=
p
k=1
x
1
(k)x
2
(k).
(Proof) The following equation is obtained from (3) and (4):
2
e
m
(k)
2
2,p
w
m
(k)
2
2,p
=
2
_
_
_
_
e
m
(k) + q
e
(k)
h
_
_
_
_
2
2,p
_
_
_
_
w
m
(k) q
e
(k)
h
_
_
_
_
2
2,p
=
2
e
m
(k)
2
2,p
w
m
(k)
2
2,p
+
2q
h
w
m
(k) + e
m
(k), e
(k)
p
. (20)
Thus, (19) is satised by using the left inequality of (17).
In regard to the input of n
(), the following inequality can be obtained from (20) and the
second inequality of (17) as follows:
w
m
(k)
2,p
e
m
(k)
2,p
, (21)
when inequality (18) is satised.
5. Sum of trapezoidal areas
The left side of inequality (18) can be expressed as a sum of trapezoidal areas.
[Lemma3] For any step p, the following equation is satised:
(p) := w
m
(k) + e
m
(k), e
(k)
p
=
1
2
p
k=1
( f (e
(k)) + f (e
(k 1)))e
(k). (22)
(Proof) The proof is clear from (15).
In general, the sum of trapezoidal areas holds the following property.
[Lemma4] If inequality (10) is satised in regard to the discretization of the control system,
the sum of trapezoidal areas becomes nonnegative for any p, that is,
(p) 0. (23)
(Proof) Since f (e
(k)) belongs to the rst and third quadrants, the area of each trapezoid
(k) :=
1
2
( f (e
(k)) + f (e
(k 1)))e
(k) (24)
is nonnegative when e(k) increases (decreases) in the rst (third) quadrant. On the other
hand, the trapezoidal area (k) is nonpositive when e(k) decreases (increases) in the rst
(third) quadrant.
Strictly speaking, when (e(k) 0 and e(k) 0) or (e(k) 0 and e(k) 0), (k) is
nonnegative for any k. On the other hand, when (e(k) 0 and e(k) 0) or (e(k)
192 Applications of Nonlinear Control
Nonlinear Phenomena and Stability Analysis for Discrete Control Systems 7
0 and e(k) 0), (k) is nonpositive for any k. Here, e(k) 0 corresponds to e
(k) =
or 0 (and e(k) 0 corresponds to e
k=1
(k). (25)
Therefore, the following result is derived based on the above. The sum of trapezoidal areas
becomes nonnegative, (p) 0, regardless of whether e(k) (and e
(k), f (e
(k))) in the rst (third) quadrant. (Here, without loss of generality, the
response of discretized point (e
(k), f (e
2
V
2
+ (U
2
+ V
2
)(1 + U)
2
+ V
2
U
2
+ V
2
, (26)
[0,
c
],
c
: cutoff frequency
when the linearized system with nominal gain K is stable. Here, () is the distorted
frequency of angular frequency and is given by
(e
jh
) = j() = j
2
h
tan
_
h
2
_
, j =
1. (27)
In addition, U() and V() are the real and the imaginary parts of KG(e
jh
), respectively.
(Proof) Based on the loop characteristic in Fig. 5, the following inequality can be given in
regard to z = e
jh
:
e
m
(z)
2,p
c
1
r
/
m
(z)
2,p
+ c
2
d
/
m
(z)
2,p
+sup
z=1
[F(, q, z)[ w
m
(z)
2,p
.
Here, r
/
m
(z) and d
/
m
(z) denote the ztransformation for the neutral points of sequences r
/
(k)
and d
/
(k), respectively. Moreover, c
1
and c
2
are positive constants.
By applying inequality (21), the following expression is obtained:
_
1 sup
z=1
[F(, q, z)[
_
e
m
(z)
2,p
c
1
r
/
m
(z)
2,p
+ c
2
d
/
m
(z)
2,p
. (28)
193 Nonlinear Phenomena and Stability Analysis for Discrete Control Systems
8 Nonlinear Control
Therefore, if the following inequality (i.e., the small gain theorem with respect to
2
gains) is
valid,
[F(, q, e
jh
)[ 1/, (29)
the sequences e
m
(k), e
m
(k), e(k) and y(k) in the feedback system are restricted in nite values
when exogenous inputs r(k), d(k) are nite and p .
By substituting (9) into inequality (29), the following is obtained:
(1 + jq())KG(e
jh
)
1 + (1 + jq())KG(e
jh
)
<
1
. (30)
From the square of both sides of inequality (30),
2
(1 + q
2
2
)(U
2
+ V
2
) < (1 + U qV)
2
+ (V + qU)
2
Then,
2
(U
2
+ V
2
) +2qV (1 + U
2
) + V
2
< 0. (31)
Consequently, as a solution of inequality (31),
<
qV +
_
q
2
2
V
2
+ (U
2
+ V
2
)(1 + U)
2
+ V
2
U
2
+ V
2
can be given.
Since inequality (26) in Theorem1 is for all (and ) considered and a certain q, the condition
is rewritten as the following maxmin problem.
[Corollary] If the following inequality is satised, the discretetime control systemwith sector
nonlinearity (2) is robust stable:
< (q
0
,
0
) = max
q
min
(q, ), (32)
when the linearized system with nominal gain is stable.
In this study, a nonconservative sufcient condition for the stability of discretetime and
discretevalue control systems is derived by applying the concept of robust stability in our
previous paper(Okuyama et al., 2002a). The stability condition is, however, not satised for
the entire area of the input of nonlinearity N(e) because of the stepwise and pointtopoint
characteristic. Even if the response seems to be asymptotic, there may remain a uctuation
(a sustained oscillation in the discrete time) or an offset. Of course, a divergent response
that reaches the sustained oscillation may occur. These responses are typical nonlinear
phenomena. The theorem (and corollary) derived here should be considered as the robust
stability condition in a global sense. In addition, it is valid based on an assumption in the
relationship between the sampling period and the system dynamics. However, this result will
be useful in designing a discrete (digital, packet transmission) control system in practice.
Naturally, the stability condition becomes that of continuoustime and continuousvalue
nonlinear control systems, when the sampling period h and the resolution approach zero.
Inequality (26) in Theorem1 corresponds to Popovs criterion for discretetime systems and
contains the circle criterion for nonlinear timevarying (discretetime) systems in a special
case. The relationship between them will be described in the next section.
194 Applications of Nonlinear Control
Nonlinear Phenomena and Stability Analysis for Discrete Control Systems 9
7. Relation to Popovs criterion
Inequality (30) can be rewritten as follows:
H(, q, e
jh
)
1 + H(, q, e
jh
)
< 1, (33)
where
H(, q, e
jh
) =
(1 + jq())KG(e
jh
)
1 + (1 )KG(e
jh
)
.
From (33), the following inequality is obtained:
2 1H(, , q, e
jh
) +1 > 0. (34)
Therefore, the following robust stability condition can be given:
1
_
1 + (1 + )KG(e
jh
) +2jq()KG(e
jh
)
1 + (1 )KG(e
jh
)
_
> 0, (35)
which is equivalent to inequality (26). When = 1 is chosen, (35) can be written as follows:
1
K
m
+1(1 + jq())G(e
jh
), (36)
where K
m
= 2K. In this case, the allowable sector of nonlinear characteristic N() is given as
0 N(e)e K
m
e
2
, e ,= 0. (37)
When h approaches zero (or is a low frequency), inequalities (36) and (37) are equivalent to
an expression of Popovs criterion for continuoustime systems.
In case of q = 0, the left side of (26) becomes the inverse the absolute value of complementary
sensitivity function T(j).
(0, ) =
_
(1 + U
2
)) + V
2
U
2
+ V
2
=
1
[T(j)[
> . (38)
On the other hand, from (35)
1
_
1 + (1 + )KG(e
jh
)
1 + (1 )KG((e
jh
)
_
> 0 (39)
is obtained. Inequalities (38) and (39) correspond to the circle criterion for nonlinear
timevarying systems.
8. Validity of Aizermans conjecture
In the following case, Theorem1 becomes equal to the robust stability condition of the
linear interval gain that corresponds to Aizermans conjecture which was extended into
discretetime systems (Okuyama et al., 1998).
[Theorem2] If the right side of (32) is satised at the saddle point,
_
(q, )
q
_
q=q
0
,=
0
= 0, (40)
195 Nonlinear Phenomena and Stability Analysis for Discrete Control Systems
10 Nonlinear Control
2
v
2
+ (U
2
+ V
2
)(1 + U)
2
+ V
2
. (41)
From (40), the following can be obtained:
(q
0
,
0
)(
0
)V(
0
) = 0. (42)
Obviously, (q, ) > 0. Moreover, since 0 <
0
< /h, (
0
) > 0 from (27). Then,
V(
0
) = 0 (43)
is obtained. Thus,
(q
0
,
0
) =
[1 + U(
0
)[
[U(
0
)[
> (44)
Inequality (44) corresponds to the stability condition which was determined for the
timeinvariant discretetime system with a linear gain, i.e., the Nyquist stability condition
for a discretetime system.
Theorem2 shows that the robust stability condition for a linear timeinvariant system (the
concept of interval set) can be applied to nonlinear discretetime control systems, when (40) is
satised. However, (32) is not always valid at the saddle point given in (40). In the following
example, it can be shown that there are counter examples of Aizermans conjecture extended
into the nonlinear discretetime systems.
9. Numerical examples
In order to verify the theoretical result, two numerical examples for discrete control systems
with saturation type nonlinearity are presented.
196 Applications of Nonlinear Control
Nonlinear Phenomena and Stability Analysis for Discrete Control Systems 11
(a)
e(k)
e(k)
(b)
Fig. 7. Time responses of e(k) and e
= (double)(int)(e/),
v = 0.4 e
+3.0 atan(0.6 e
),
v
= (double)(int)(v/),
where (int) and (double) denote the conversion into an integral number (a rounddown
discretization) and the reconversion into a doubleprecision real number, respectively.
When choosing the threshold = 2.0, the sectorial area of the stepwise (pointtopoint)
nonlinearity for [e[ < 35.0 can be determined as [0.5, 1.5] drawn by dotted lines in the
gure. In this example, the sampling period is chosen as h = 0.1. From (26) and (32), the
maxmin value can be calculated as follows:
max
q
(q,
0
) = (q
0
,
0
) = 0.49,
when the nominal gain K = 1.0. Hence, < 0.49 and the stable area is determined as
[0.51, 1.49]. Obviously, this sector contains the area bounded by the dotted lines. Thus, the
discrete control system is stable in a global sense.
The stability condition for linear gain K can be calculated as 0 < K < 1.5 when the
sampling period is h = 0.1. In this example, Aizermans conjecture for discretetime system is
satised. Figures 7 (a) and (b) show time responses e(k), e
(k)
(a)
e
(k)
(b)
Fig. 9. Backward difference e
(k), e
(k)) of the discrete nonlinear control system when the reference inputs are r =
1.0, 2.0, 3.0. Figures 8 (a) and (b) show those responses when r = 5.0, 7.0, 9.0. Although
the responses contain sustained oscillations, they do not exceed the threshold = 2.0. The
input and the output of the nonlinearity lie in a parallelogram shown in Fig. 6. The robust
stability in a global sense is guaranteed for all the reference inputs r. The above behavior can
be estimated fromthe intersections of the highest gain of the sector and the stepwise nonlinear
characteristic. Obviously, discretevalues (1.0, 2.0) and (1.0, 2.0) lie in the outside of the
stable sector.
Figures 9 (a) and (b) show the traces of backward difference e
Fig. 10. Discretized nonlinear characteristic and stable sector for Example2.
(a)
e(k)
e(k)
(b)
Fig. 11. Time responses of e(k) and e
(k), e
(k)
(a)
e
(k)
(b)
Fig. 13. Backward difference e